Skip to content

armgit / Pulse

Public Imported from GitHub
0

Real-time monitoring dashboard for Proxmox VE, PBS, Docker, Kubernetes, TrueNAS and vSphere. Self-hosted, with alerts for failed backups, full disks, unhealthy containers and offline nodes.

763 KB
Raw Blame

Frontend Primitives Contract

Contract Metadata

{
  "subsystem_id": "frontend-primitives",
  "lane": "L8",
  "contract_file": "docs/release-control/v6/internal/subsystems/frontend-primitives.md",
  "status_file": "docs/release-control/v6/internal/status.json",
  "registry_file": "docs/release-control/v6/internal/subsystems/registry.json",
  "dependency_subsystem_ids": [
    "agent-lifecycle",
    "api-contracts",
    "cloud-paid",
    "storage-recovery"
  ]
}

Purpose

Backups bars use the metric bar colours (2 November 2026)

The Backup health rows (ProxmoxBackupsCoverageStrip) and the Backup jobs chart (ProxmoxBackupJobsCard) fill their bars with bg-metric-normal-bg, bg-metric-warning-bg and bg-metric-critical-bg, the tokens every getMetricColorClass bar uses. They had used Tailwind emerald, amber and red at their own opacities, so the Backups page green visibly differed from the Overview CPU, memory and disk bars. Status dots and count pills keep their own colours; only bar fills moved.

Drawer History keeps live readings when no chart can be drawn (2 November 2026)

When the first history read fails, or the selected range is beyond the plan, GuestDrawerHistory shows a "Latest readings" strip in place of the charts. It reads the same currentMetrics and deferredMetrics as the chart readouts, through one getLiveReading helper, so retained or freshness-unknown memory keeps the qualifier and explanation from getMemoryObservationPresentation. The strip draws no plot, timestamp or inspection. A final 401/403 still withdraws the whole panel, readings included, and a successful read restores the charts. PBS servers depend on this: their host CPU and memory are read in the drawer, not the Backups table. GuestDrawerHistory.refresh.test.tsx and ResourceDetailDrawer.history.test.tsx (a PBS drawer with a rejected history read) cover it.

Compact toggles mark selection by colour, and icon labels stay on one line (2 November 2026)

The compact FilterButtonGroup variant (View preferences, type and status chips, the history range) marked its default-tone selection only with bg-surface on a bg-surface-hover track. In dark mode that fill is darker than the track, so the selected option read as unselected. The default active tone in filterButtonGroupModel.ts now carries blue text and a blue ring (text-blue-700 dark:text-blue-300 ring-blue-200 dark:ring-blue-800), as the default and segmented variants already do; explicitly toned options keep their own colours. A visualLabel that contains an icon wraps in an inline-flex items-center gap-1 span in FilterButtonGroup.tsx, because preflight makes an svg a block box and stacked Layout and Metrics icons above their text. FilterButtonGroup.test.tsx covers the selected tone and the icon row. No option, value, keyboard or count behaviour changes.

Audit viewer state belongs to its reader and accepted page

The existing Audit Log panel receives live principal, non-secret token/scope identity and explicit audit-read capability from the Settings registry. Reader, organisation, entitlement or lifetime changes cancel work and withdraw events, counts, signature results and batch feedback. An equal-reader status refresh keeps its pending operation. A fresh inventory read withdraws the previous page and all its signature work before awaiting a replacement; an old result cannot verify a new event with the same ID. Only current accepted row objects admit verification. Reads and signature checks bind the expected organisation with cancellation and no automatic retries. Cancel ends local verification immediately and Resume is an explicit action; neither an old batch nor its completion may change a new batch's progress or publish feedback.

Failed, refused and malformed reads are unavailable, not empty or verified. Verification refusal withdraws the page and requires a new read; other check failures remain retryable on the current row. Inventory validation accepts the producer's omitted optional strings. AuditLogReaderLifetime, existing audit request/branch tests and registry tests run in the normal frontend suite. Desktop/narrow actual-App proof covers the existing route, including late responses and usable Refresh/Verify/Cancel/Resume. This changes no audit source, server permissions, signing semantics, retention or installed recovery contract.

Token reveals keep content and copy feedback with the admitted value

The existing token reveal consumer scrolls within the bounded shared dialog, including short, narrow and enlarged-text viewports. The title/source and Close control use separate flow space; literal notes, labels, hints and token values wrap without hiding Copy or Dismiss. Shared dialog focus and dismissal ownership are unchanged.

Replacing the reveal updates the actual value, record, hint, source and note together; the consumer never retains a snapshot from the previous reveal. Only one deliberate clipboard request runs per reveal at a time. Pending feedback is distinct from success; a result is reported only while its originating reveal is current and mounted, even when a later reveal has identical token bytes. Replacement/dismissal/unmount invalidate feedback and its reset timer, not an already-requested system clipboard write. Token generation, reveal admission, authorisation and API/storage/clipboard permissions are unchanged. Normal-suite TokenRevealDialog.test.tsx covers current success/failure/retry, pending dedupe, replacement, dismissal, unmount and timer isolation. Source-bound browser proof must exercise the real store/dialog/clipboard helper at desktop and narrow/short viewports, with native scroll, keyboard/touch, deferred results and inspected pixels.

Backup views share bounded owned-focus restoration

Coverage posture sorting and By date regrouping use the same existing backups presentation helper for focus lost during a retained native DOM move. It restores only the same connected target still inside its table and cancels after any intervening focus choice, even if that choice is later blurred. By date's earlier restoration did not cancel that latter choice; its normal-suite date-group regression now covers it. Genuine retirement still uses the existing surface fallback. No sorting, date validity, collection, protection or access rule changes. The consolidation reduces overlapping handling. Its exact build still exceeds the August Proxmox reference budget by 0.07 KiB; the final candidate declares a bounded 256-byte Proxmox reference allowance (80,595 → 80,851), not a measured baseline reset. The 15% threshold, every other chunk and total limits remain unchanged. Normal Proxmox/shared suites and exact production checks verify it. Desktop/narrow actual Backups proof includes both views and source-bound pixels.

Coverage owns focus across native keyed-row movement

Stable keyed row components alone do not preserve native browser focus when a posture update changes the sorted DOM order. The existing Coverage consumer captures only a target inside its own table before reconciliation and restores the same connected target without scrolling if native movement leaves BODY focused. Any intervening focus choice cancels that restoration, including a choice later blurred to BODY. Removed rows and departed tables are not refocused; external search/header controls and existing retirement fallbacks keep ownership. This does not change the shared renderer, sort order, row budgets, read cadence, posture meaning or access. ProxmoxCoverageTable.test.tsx exercises wide/narrow VM/LXC moves, same-node detail/disclosure continuity and no-steal/retirement controls. Actual Backups desktop/narrow pixels and interactions check poll-driven movement against a source-bound parent control.

Coverage keeps provider provenance separate from task state

The existing Coverage drawer uses the shared wrapping detail-section pattern for each provider's scope, evidence source and Reported outcome. Known PBS/PVE inventory sources are named; unknown sources stay literal text. The summary is not labelled a job or collection result because the rollup can combine sources of different ages. A visible qualifier distinguishes provider evidence from actual task and restore-point states without changing their semantics or shared primitives. ProxmoxCoverageTable.test.tsx and ProxmoxBackupsTable.test.tsx cover source meaning, unknown/literal values, independent states, polling and focus. Actual desktop/narrow Backups pixels must confirm readable wrapping.

Coverage dates are inspectable without hover

Open Coverage evidence retains each point’s exact reported date as wrapping literal text beside its identifier and metadata, including seconds, fractions and timezone offsets. Missing dates say “Not reported”. Invalid and future raw dates remain inspectable without becoming valid ages or ordered evidence. The existing Unknown age and uncertainty notices still apply. This works in both the Details column and its folded phone projection; it adds no read or recovery operation and does not change completion, posture, chronology or preview rules.

Normal-suite coverageRestorePointIdentity.test.tsx exercises parsed PVE/PBS inventory for all three point kinds, valid/invalid/future/missing/literal dates, responsive projections, same-row polling and namesake isolation. Desktop and narrow browser proof inspects the real Backups consumer and its updated dates. The canonical bytes formatter supplies the existing Unknown size fallback in both projections; no byte admission or sorting rule changes.

Coverage restore points retain their identifiers without hover

The existing open Coverage evidence rows show each point's full identifier and metadata, including an archive's full volume identifier rather than only its format/hover title. When the Details column is hidden, those same facts fold beside the source. Literal text wraps at every width; it is not interpreted as markup. Same-time points remain distinguishable without changing inventory identity, preview selection/order/limits, requests, counts, posture or state. Ordinary polling updates the same open rows and preserves disclosure focus.

Normal-suite coverageRestorePointIdentity.test.tsx consumes parsed PBS/PVE inventory with tied dates for all three source kinds, literal markup, long volume identifiers, an incomplete point and namesake guests. It covers responsive projections and same-panel polling. Desktop/narrow browser proof exercises the real Backups consumer with readable, inspected identifier and metadata pixels. This is safer evidence inspection, not a restored guest or installed recovery.

Backup size uncertainty survives responsive disclosure

The Backups consumer uses the existing byte formatter only for a positive reported size. Its row and open point explicitly label unavailable sizes rather than hiding them; Coverage folds the same per-point fact beside its source when the size column is hidden. Wrapping explanatory text keeps unknown distinct from an empty point without adding a shared primitive or changing byte units, layout breakpoints, control ownership or focus. Storage/recovery owns the size admission and ordering. Normal Proxmox consumer tests and desktop/narrow browser inspection cover the actual disclosure, including same-point polling.

Selected PBS datastore remains inspectable without hover

The existing PBS row disclosure is named by backup reader and datastore, rather than giving every store on one server the same control label. Single-line scan cells and responsive widths are unchanged. Its open detail presents the full, wrapping reader/store identities and admitted capacity bytes/percentage and inventory count, including when those columns are hidden or truncated. The shared platform detail row accepts optional row context inside the same native inline shell; this does not replace the canonical resource or its History target.

Loading and unavailable inventory remain nonnumeric; accepted empty inventory is zero. Unavailable/unknown capacity never becomes zero, even with a Healthy connection. A server without datastore data has neither a made-up store nor a count. Counts describe loaded entries, including in-progress/unverified entries, not backup protection. These independent evidence meanings, correlation, read cadence, retry, API, auth and protection policy are unchanged.

Normal-suite backupServerDatastoreAccess.test.tsx exercises long namesakes, responsive projections, parsed inventory loading/failure/retry/polling, capacity admission, no-store data, live reordering and retained detail/focus. The registered ProxmoxBackupsTable.test.tsx verifies the selected consumer; shared detail-row and platform-layout tests preserve the separate context/host owners. Browser proof exercises actual keyboard/touch disclosure, open-record refresh and complete wrapping identifiers at desktop/narrow widths with pixel inspection. This is safer evidence inspection, not original guest or installed recovery.

Backup-server read states remain legible at intermediate widths

The existing PBS server table reserves enough count-column width for the full Loading and Unavailable labels, not only numeric counts, at basic, operational and expanded container widths. The shared 30% basic identity anchor, visible columns, 32px scan rhythm, independent connection/capacity evidence and full layout remain unchanged. Datastore names keep their existing truncation, title and detail access; no identity, count, polling, read/retry or protection rule changes. An unread inventory still never establishes a real zero.

Normal-suite proxmoxBackupsTablePresentation.test.ts verifies width allocation and complete column totals; the registered ProxmoxBackupsTable.test.tsx exercises actual loading/failure/accepted-empty cells with independent healthy capacity. Browser proof must inspect the consuming table at desktop, phone and intermediate container boundaries, including loading, failure, retry, real zero, ordinary polling and native detail interactions. This is readable inspection, not installed or original guest recovery.

PBS server counts belong to the exact backup reader and datastore

The existing Backups server count matches PBSBackup.instance to the resource's configured connection name and matches the exact datastore name. Service IDs (pbs- plus the connection name), hostnames and display labels are not backup reader aliases; an empty tuple must not borrow another connection's inventory. Both identities retain their positions, punctuation and case without a joined separator key. Counts still include all admitted inventory entries and namespaces, not only completed or verified points. Loading/unavailable inventory remains non-numeric; independently observed capacity and reachability remain unchanged.

Normal-suite backupServerCountEvidence.test.tsx consumes parsed inventory and checks prefix-alias and delimiter collisions, case, namespaces, valid zero and existing all-entry semantics. Its real table consumer checks ordinary polling, failed reads, retry and detail continuity without extra requests. Desktop and narrow browser proof exercises the same producer through the inventory consumer with native interactions and inspected pixels. No API, provider, guest, polling, auth, resource identity or protection-policy change, nor installed recovery claim.

Coverage duration requires observed task timing

The existing Coverage task disclosure derives a duration only from positive, strict RFC3339 endpoints at or before the inventory observation time, with the end after the start and no reported running status. Future/invalid/nonpositive or reversed endpoints, or a running task carrying an end time, show Unknown and an explanation, not a numeric completed duration or Not reported. An absent end remains Not reported; a valid positive subsecond interval may round to 0s. Valid offsets are compared as instants. Task-reported status, errors, identity, latest-task ordering, protection and restore evidence remain independent and unchanged. Ordinary polling can withdraw and restore duration within the same inspected task without replacing the panel or taking focus.

Normal-suite taskDurationEvidence.test.ts consumes parsed raw inventory and covers adverse endpoints/status plus valid, missing, offset and zero controls. backupTaskEvidence.test.tsx exercises those states through the actual inventory consumer and ordinary polling. Desktop/narrow browser interactions and inspected pixels cover the changed producer-to-disclosure path. No new requests, provider operation, collection policy or native guest recovery is implied.

Replication sync chronology stays unknown when unreadable

The existing Replication table and disclosure validate sync instants before passing them to shared age formatting or building a countdown. Impossible dates, timezone-less strings, out-of-range numbers and future last-sync values do not become repaired dates, overdue schedules or just-now successes. Unknown remains visible and distinct from unreported/zero values; the latter are not proof that a job never ran. Narrow cells use a question mark with a full accessible name/title, while Schedule details and the summary retain full wording. The existing Unix companion precedence, supported numeric forms, provider status, shared aging clock, disclosure controls and read/retry behaviour remain intact. ProxmoxReplicationTable.test.tsx consumes parsed raw inventory for unknown, future, unreported and valid controls; desktop/narrow browser interactions and pixel inspection cover the same producer-to-consumer path. This changes no shared formatter or primitive and does not establish native job/guest recovery.

Withdrawn backup evidence retains keyboard position

When polling removes a focused restore-point disclosure/detail, filtered Coverage row, activity chart or the whole inventory body, focus moves to an existing inventory retry if a source failed, otherwise the active Backups view. Kiosk mode falls back to the named inspection region. The fallback only takes focus owned by the retiring content, after removal, while the inspection is still connected and no other control has acquired focus; it never scrolls the page. A normal detail close still returns to that same point's surviving disclosure. Other-source readable evidence, query/filter state, native controls and the existing retry cadence remain unchanged; no stale point is retained to preserve focus. Filtered charts and unmatched-evidence disclosures use their own retiring owners. InlineDetailTableRow composes its internal native row ref with the consumer's ref; registering a consumer cleanup must not disable its responsive span or normal close-focus return. Its normal unit suite verifies both with a consumer ref present. No new shared component or inspection wrapper is introduced.

Normal-suite backupInventoryFocus.test.tsx consumes parsed raw PVE/PBS responses and covers both sources, summary/detail retirement, Coverage lookup, chart and whole-body removal, accepted-empty polls, normal close, moved focus, kiosk and unmounting. ProxmoxBackupsTable.test.tsx also covers a later polling-failure notice after a fulfilled read, through pending retry and complete recovery. Actual production-App desktop/narrow browser proof exercises native keyboard focus and read-failure/retry through the real consumer, with inspected pixels. This is accessible inspection, not guest thaw, restore or installed recovery. No provider operation, route, auth or policy changes.

The existing backup text search matches admitted native restore-point identifiers in both By date and Coverage, and full archive metadata in Coverage. This is a consumer haystack correction: the shared FilterBar/SearchInput, literal search grammar, keyboard and touch controls, and exact scope-filter keys are unchanged. Normal ProxmoxBackupsTable.test.tsx and actual-App desktop/narrow proof cover input and cross-view retention without additional provider reads.

Summary buttons activate once per Space press

Native summary disclosure and scope buttons retain their existing keydown-owned Space action. They cancel the corresponding keyup default as well: mobile Chromium can otherwise synthesize a second click and undo the activation. Enter, Escape, pointer actions, expanded/pressed semantics and focus ownership stay unchanged. Normal-suite SummaryRowActionButton.test.tsx checks both button kinds and that Enter keyup is not suppressed. Browser proof must exercise actual App disclosure Space/Enter, touch and focus at desktop and narrow widths.

TrueNAS protection consumers do not repair malformed instants

The Protection row, completion sorts and inline Started/Completed fields use the shared explicit-zone, calendar-valid backup instant rule. Invalid reported values remain Unknown, with the literal evidence retained in details rather than a browser-repaired date. Unknowns have no chronological rank; outcome priority, absent/active completion handling, identity and policy remain unchanged. TrueNASProtectionTable.test.tsx drives the query/transport normalizer into these consumers with malformed and valid-offset controls; desktop/narrow browser pixels and interactions remain required. This is not native protection or guest recovery.

Missing Workloads completion times do not establish absence

The workload badge, compact shield, hover/dialog and full drawer treat absent completion-time sentinels as missing current-inventory evidence, not proof that no backup exists or that a running backup is the first. The legacy never key stays internal; visible absence is cautionary Unknown and “No completed backup reported”, with no invented date or age-rule link. The disclosure states that missing data does not mean no backups exist. Known valid ages, invalid/future time reasons, blue running activity, platform/template gating and independent guest-read/thaw precautions are unchanged. No read, timer, policy, identity, guest, backup or restore operation is added.

backupTimestampFlow.test.tsx drives missing/empty/Go-zero observations through the real REST and owning-snapshot mapper into badge, shield and full drawer for VM/LXC, including activity, recovery/recurrence and retained dialog/focus. Existing cell, drawer, shared-clock and guest-row tests pin the accessible wording, warning tone and separate safety evidence. Desktop/narrow browser proof must inspect real production rows and their consuming disclosures, not just helpers. This is truthful inspection, not original protection causality, native thaw, restore readiness, installed recovery or release availability.

Backup read recovery retains keyboard position

PVE, PBS and protection-posture retries remain focusable while pending, report aria-disabled, and reject duplicate activation. A failed retry retains its control. When a recovered notice or whole-inventory error is removed, focus owned by that content moves to the existing active Backups view link, or the named inspection region when kiosk mode hides navigation. Recovery never moves focus the operator has taken elsewhere, or restores it after leaving the surface. Independent inventory, server health, assessment withdrawal, read cadence, filters, guest operations and protection policy are unchanged.

ProxmoxBackupsTable.test.tsx exercises pending/failed/successful retries, duplicate activation, partial whole-inventory recovery, moved focus, kiosk fallback and disposal. Desktop and narrow browser proof must exercise the real HTTP clients, keyboard input, ordinary polling and consuming pixels. This is inspection accessibility, not original guest or installed recovery.

Runtime alert controls describe only supported scope (28 October 2026)

Container Runtime rows expose their existing Offline Alerts control, not an all-alerts switch or metric bulk editor that cannot mutate runtime settings. Visible guidance explains that connectivity settings do not mute container or machine alerts, and directs operators to existing per-container Alerts switches and saving; newly added containers need their own setting. No single-runtime child mute, inheritance, policy evaluation, persistence or machine-wide mute is introduced. Existing global-disabled connectivity precedence remains intact.

The shared resource table can omit the Alerts column for this connectivity-only surface, including row, group, default, delay and empty-state column counts; other tables retain their existing controls by default. Phone cards honour the same choice. ResourceTable.test.tsx and ThresholdsTable.test.tsx verify the omitted controls, remaining connectivity action, independent container drafts and preserved machine/other-runtime settings. Desktop and narrow browser proof must join the real projection, mutation hooks and consuming table, not an Off label unsupported by the producer. This is truthful control presentation, not fulfilment of discussion #3432/#2107 whole-runtime-and-children demand, recipient proof, installed recovery or release availability.

Missing-filesystem help separates real capacity from container layers (28 October 2026)

The existing VM disk guide and Troubleshooting entry distinguish an attached virtual disk, a real local filesystem under a container-data directory and an intentionally excluded layer or remote mount. The version-specific v6.6.0 path filter limitation is not an agent-installation diagnosis or a claim of repair. Backup inclusion stays separate from capacity collection; deliberately excluded disks stay excluded. Existing guest or already-linked-agent observations offer an interim capacity view, not API recovery or complete monitoring.

missingFilesystemGuide.test.ts verifies the production Markdown renderer, entry/target links, exact shipped mirrors, narrow-friendly prose instead of a new table, source/time/privacy distinctions and the absence of new commands. No remount, data move, service restart, new agent, guest probe or backup change is prescribed. Backup safety retains precedence. Source-bound desktop/narrow Docs interactions and inspected captures remain required; this is Help maintenance, not a filesystem-admission fix, installed recovery or publication.

Coverage evidence remains inspectable without a pointer

The bounded Coverage preview includes undated source/state witnesses and a plain count of dates which cannot be ordered. Its native expansion button adds eight evidence entries at a time without another read. The same control keeps keyboard focus and reports when all entries are shown, including after polling; closing or switching workload resets the preview. Provider text wraps literally at narrow widths. coverageEvidenceInspection.test.tsx and ProxmoxCoverageTable.test.tsx check real and direct consumers, preview bounds, literal origins, scope, focus, recovery/recurrence and reset. Desktop keyboard and narrow touch proof must inspect the production table and actual pixels. This does not change shared primitives, protection policy or guest operations.

Backup age and completion do not normalise malformed strings

The backup badge, shield disclosure and guest drawer share Backups inventory's strict calendar/zone parser. A present impossible, unzoned or date-only time is Unknown, never a green completed backup or an invented exact UTC completion. Numeric/Date inputs and valid zoned instants retain policy-relative age, current activity and existing absence/future semantics. backupTimestamp.test.ts and backupTimestampFlow.test.tsx cover direct formatting and REST/snapshot production consumers, recovery/recurrence, details and focus. Desktop/narrow browser proof must exercise these changed producers through the real consumers. See storage-recovery's backup completion rule; no restore/thaw claim is added.

Latest backup task evidence stays separate from recovery

Coverage’s existing expanded detail shows the scoped latest PVE task’s reported status, start, duration, full literal error and task ID. Long/multiline provider text wraps without markup interpretation. An unreadable start says Unknown, a missing duration says Not reported, and a failed task without error text says Not reported. No task means no task section. Ordinary polling updates and removes stale task details in the same keyed panel without losing focus. The detail explicitly does not verify a restore or guest responsiveness. That qualification wraps at narrow widths instead of inheriting cell nowrap.

backupTaskEvidence.test.tsx exercises the HTTP inventory parser, recovery model and production Coverage consumer at desktop/narrow layouts, namesake connection isolation, literal text, missing/invalid evidence, running/OK tasks and same-panel polling/removal. Desktop keyboard and narrow touch browser proof must inspect these existing details. No new read, route, task action, matching rule, protection policy, backup/restore or guest operation is added.

Backup workload scope stays visible independently of its reader

Coverage retains the guest's PVE connection in its identity cell at every container width. When Target ID is hidden, type and VMID remain beside the name. Native disclosure names include the workload label and PVE connection; the existing Coverage and By date Workload details show the full wrapped connection, separate from the restore point's backup-reader Connection. A missing VM/LXC connection stays Not reported, including unmatched PBS evidence; PBS readers, namespaces and nodes never supply it. Host backups imply no PVE connection. Displayed scope does not change identity matching, resource/artifact keys, owner-link IDs, search/filter behaviour, completion, protection policy, verification, access or guest/recovery operations.

The normal-suite backupWorkloadIdentity.test.tsx exercises inventory parser, recovery model and both table consumers, five layouts, duplicate names/IDs/nodes, unknown/unmatched/host scope, distinct PBS origins, exact owner links, literal text and same-key live updates with disclosure focus. Existing windowed-table proof remains applicable; the Coverage row-height estimate accounts for its second identity line. Browser proof must inspect real desktop/narrow consumers, full long/unknown details and keyboard/touch interactions. This does not establish original multi-surface identity, collector, guest or restore recovery.

Restore evidence names its source connection

The existing By date source cell and Coverage restore-evidence source cell show the backup reader's connection alongside its provider kind. Repository, namespace, guest name and node are not substitutes for that connection: two independent sources may share all of those labels. The list keeps the source inside its column; the existing restore-point detail names the full Connection separately from Location, and Coverage wraps the full connection even when its Location column is hidden on phones. Missing source identity stays explicit as Unknown connection / Not reported, never inferred from the workload. Both existing text searches match the displayed connection (including Unknown connection), without changing the exact location filter or broadening its key.

backupSourceConnection.test.tsx exercises the actual API/model/table path for PBS, PVE archives and guest snapshots, same-named sources, missing origins, literal text, unchanged posture and same-point polling with retained detail and focus. Desktop/narrow consuming browser proof checks keyboard/touch access and readable long source names. Exact filter keys, IDs, namespace, chronology, completion, verification and server-owned protection policy are unchanged. This is source-selection clarity, not an installed restore or guest recovery.

Backup retention never substitutes for verification

The existing restore-point State cell and drawer status prioritise running or failed completion and then the reported verification result. A point kept when pruning still says Unverified or Verified as reported. Retention without a verification verdict says Kept with an informational tone, never Protected. The existing detail names Prune protection separately and says Not reported when backup verification is unavailable. Guest snapshots retain their existing informational state. The shared Coverage evidence uses the same state wording. Canonical workload protection, freshness, policy and inventory counts are not inferred from prune retention or an individual verified file.

The normal-suite backupRetentionPresentation.test.tsx exercises the real inventory consumer, independent assessment, explicit/missing verification, search compatibility and same-point polling with an open drawer and focus. Shared badge and drawer tests retain completion precedence and reactive tones. Desktop/narrow consuming browser proof checks the existing rows and details. No provider data, route, access, retention operation or guest recovery changes.

Backup filters also scope activity

The existing Proxmox By date chart and date warning consume the same feature-owned filtered inventory as the table, before the selected-day facet. Changing source/verification, search, node, type or exact location updates both counts and volume without a new read. Selecting a day narrows only the list and leaves other scoped days selectable; clearing restores the scope. An empty or unresolved scope does not advertise estate-wide activity. The page explains this relationship beside the existing chart. Aggregate backup-health counts remain server-owned and estate-wide.

The registered ProxmoxBackupsTable.test.tsx scope cases and source-bound desktop/narrow browser journeys cover the shared filter rail, keyboard/touch day selection, metric switching, empty matches and recovery. No shared-filter, route, navigation, access or backup operation changes.

Backup health reason rows carry a trailing chevron (2 November 2026)

ProxmoxBackupsCoverageStrip's row grid gains a last track for a lucide ChevronRightIcon on the reason rows that are buttons, flush right on desktop and phone, so a filtering row looks clickable without adding height. Rows that open nothing leave the track empty, so every row's columns still line up.

Backup servers rows group under a header per server (2 November 2026)

ProxmoxBackupServersTable renders one getGroupedTableRowClass header row per PBS resource, spanning every column, before that resource's datastore rows: a connection StatusDot, ResourceNameWithWebInterfaceLink for the server, and a datastore count in GROUPED_TABLE_ROW_META_CLASS, omitted when the server has no datastore rows. Datastore rows show only the datastore name; the muted server name and per-row web link are gone. groupBackupServerRows builds the items, keyed server: and datastore: so an item never changes kind, and the header carries plain values only so its windowed store never aliases a datastore row's resource snapshot. Rows stay single-line. The table's PlatformTableShell title reads "Backup servers" with the server count, like the Overview "Nodes" strip. Connection health renders only on the server header (red label when the server is not reachable); status is a compact-layout column ("State") that answers for the datastore with a fill forecast, a job problem, its job result, or "Unreachable".

Backup servers rows lead with the datastore (2 November 2026)

getBackupServerColumns drops the server and separate datastore columns for one datastore identity column (kind: 'name') that keeps the shared identity anchor (40% on phones, 30% in the basic band). ProxmoxBackupServersTable renders the datastore name in it, unshrinking up to the cell width, followed by ResourceNameWithWebInterfaceLink for the server in 10px muted text, which truncates first. The status cell renders on every row.

Backups By date carries no activity chart (2 November 2026)

The restore points no longer render BackupActivityChart; the component and its presentation module are deleted. Backup jobs is the page's one daily bar chart, so two look-alike bar charts never stack on one page. Earlier notes here about the chart's tooltip portal and day-control focus are historical.

Backups source labels keep "snapshot" for guest snapshots (3 November 2026)

proxmoxBackupSourcePresentation names PBS items "PBS backup(s)" in every label (filter, detail, fallback, timeline) and gives a PBS item with no verification result the state "Backup"; the guest snapshot source keeps "Snapshot". One word, one meaning on the page.

Backup view tabs head the restore points, not the page (2 November 2026)

On the Backups page the shared PlatformSectionTabs ("Backup views": By date, Coverage) now sit at the top of the restore-point block, below the Backup health and Backup jobs cards and the Backup servers table, so a tab reads as switching the list it heads rather than the whole page. The tabs keep their nav landmark and active-route state, so recovery focus that lands on the active view still finds it inside the inspection region.

Compact Coverage columns give the name the slack (3 November 2026)

ProxmoxCoverageTable takes identityRows (every row before filters) to find namesakes, and in the compact layout shows the VM id and connection beside the name only for them. In the compact layout, once the table knows its width, the cells beside the name take their measured needs at 320px (posture 80px for "Unknown" with its dot, age 52px for an "Unknown" age, job 34px for its header) as shares of that width, and the name takes the rest; without a width the shared 40% identity share still applies. At 360px and up a 16-character guest name fits whole. Rows stay 32px and single-line.

PlatformResourceDetailTableRow can collapse its resource inspector (3 November 2026)

PlatformResourceDetailTableRow takes an optional collapseResource (title, summary, showLabel, hideLabel). When set, the row's context answers the row and the canonical ResourceDetailDrawer sits inside a ResourceDetailDrawerSupportDisclosure, mounted only when opened, so its history reads start on demand. It starts collapsed every time the row opens, and the inspector gets no onClose, because "Hide" is its close and the row's own heading closes the row. Without it every caller is unchanged. The Backups datastore drawer is the one user: its context carries the canonical ObjectDrawerHeader and DrawerSubjectHeading, so the drawer keeps a subject heading and a close control while the server inspector is collapsed.

Backup servers table drops its host columns (2 November 2026)

ProxmoxBackupServersTable no longer renders Version, CPU, Memory or Uptime columns; BACKUP_SERVER_COLUMNS is server, status, datastore, used, jobs, backups and dedup, and every layout's weights were redistributed so Used, the longest value (percentage plus the used/total pair or the forecast pill), keeps the widest datastore share. The table therefore no longer consumes formatPlatformTableUptimeValue, and the shared-template registry drops it from that rule's required consumers. The memory cell's multi-line "Last known" qualifier went with it, so every row is now single-line by construction.

Backup health lists datastore job failures in their own rows (3 November 2026)

ProxmoxBackupsCoverageStrip takes datastoreProblems and renders them as a separate "Datastores with failed jobs" list under the guest rows, on the same row grid: amber dot, "Sync failed", a datastore-count pill, and the datastore names where a reason row's share bar sits, since a datastore is not a share of rated guests. Rows stay 32px and single-line; names truncate on phones and the accessible name and title carry them in full. The list renders even when no guests are rated.

Datastore fill forecast stays on one row line (2 November 2026)

ProxmoxBackupServersTable shows a PBS datastore's fill forecast as the shared count-pill style in the Used cell, replacing the secondary used/total bytes so the row keeps its single line. In the compact layout the pill would crowd the percentage, so the forecast moves to the State cell as tinted text, the slot the compact layout already gives a job problem. Its label comes from describeStorageAlertHeadline, the Storage page's wording, and the drawer's DetailSectionTable gains a toned "Fill forecast" row. ProxmoxBackupServersTable.forecast.test.tsx and proxmoxBackupServerForecast.test.ts cover both layouts, the drawer, live alert changes, the unreadable store, the cross-connection id collision and the job-problem precedence.

Backups view tabs lead with Coverage (3 November 2026)

The Backups view tabs now list Coverage before By date, because Coverage is the default view. They keep the same route-backed links and aria-current. The Backup health reasons "No backup history" and "Backup source reported a failure" become "Backup status unknown" and "Last backup failed"; each still shares one 32px row with its count and bar at phone width.

Backup jobs header names the pills' window at every width (3 November 2026)

ProxmoxBackupJobsCard drops the "14 days" title suffix and always labels the header pills with their window: "Last 24h" from sm up, "24h" below, so a 375px header stays one 21px line with three pills. The incomplete-window tag reads "incomplete data".

Backup jobs card composes the overview primitives (2 November 2026)

ProxmoxBackupJobsCard is a TableCard with a TableCardHeader strip, like Backup health: last-24-hour outcome count pills in the header actions (they wrap rather than clip, and failed and warning guests are named in their own pill's title) and one stacked bar per calendar day. A day Pulse did not fully observe keeps a dashed outline and says so in its accessible text; it is never drawn as a clean empty day. Unavailable history replaces the pills with "History unavailable" rather than zero counts.

ProxmoxBackupsTable takes the card through its summaryAside slot. The page surface owns that read, so the table's inventory reads stay PVE plus PBS and the slot renders identically in both route-backed views. The summary grid puts the card beside Backup health from xl only while it renders and stacks it on narrower screens. ProxmoxBackupJobsCard.test.tsx and proxmoxBackupJobs.test.ts cover pill order, failure names, outlined days and unavailable or malformed history.

Backup activity uses calendar-day chart slots

The Proxmox backup activity producer supplies one slot per consecutive local date, including across daylight-saving and half-hour clock changes. The chart's labels and selectable day keys share that local-date basis; elapsed 24-hour steps cannot establish a calendar boundary. Count and volume retain identical slots, while future instants and unknown dates remain excluded. The registered activity presentation tests and source-bound desktop/narrow consuming browser journey cover date selection, metric changes and the matching backup details. No shared table, scrolling, query, policy or recovery operation changes.

Backup-location choices preserve domain identity

The existing shared filter rail renders source-qualified backup-location labels while its feature-owned values preserve the exact source/connection/ repository tuple. Labels are not identifiers. The value persists between By date and Coverage and across reloads; new choices never write the old lossy colon-delimited value. An unambiguous legacy saved value displays the resolved current option. An ambiguous or missing saved value keeps a removable "Choose location" chip and an accessible notice, not a raw URL value, another repository or a silently broadened list. The existing chip menu can select a specific current location or clear the filter without another API read.

The registered location/model/consuming tests cover identity, compatibility, empty scoped results, navigation, details and recovery through the shared rail. Source-bound desktop keyboard and narrow touch proof checks actual fit and interactions. No shared-filter implementation, route, access or backup operation changes; server-owned health remains an aggregate assessment, not a per-location recalculation.

Drawer History readout keeps its size while scrubbing (26 October 2026)

Scrubbing a drawer History chart changes the readout's text, never its size. The readout time and each series value reserve, in two invisible pseudo-elements, the widest texts the loaded range can put there (compared by shape under tabular-nums), so the legend wraps by panel width alone and the plot no longer jumps up and down under the pointer. A bucket's peak moved off the legend onto the plot: an aria-hidden label beside the cursor at the spike's height, spaced from other series' labels and flipped leftward past the middle. The slider's value text still states each reading's average and range. GuestDrawerHistory.inspection.test.tsx pins equal reserves at rest and while inspecting, and the label's place on the plot. A scripted mouse scrub of the production build at 1024 and 1440 px went from up to 80 layout changes per sweep (20 px jumps of the plot and of everything below the chart grid) to none.

Detail sections follow their own available width

DetailSectionTable owns a named inline-size container. Its existing card presentation activates only when that container has at least 64rem, not when the viewport does. Narrow drawers and resizable panes retain the compact single bordered table even on desktop, so a 7rem label track cannot consume a whole card and collapse its value to zero width. The nested scroll shell must not become the owner of the named layout query. Balanced integral card spans, rich values, progress, stacked rows, wrapping opt-ins, test hooks and the public component interface are unchanged. The registered DetailSectionTable.test.tsx and shared guardrails cover this owner and existing row/card structures; source-bound browser proof must establish nonzero readable values at narrow panel and phone widths and the unchanged cards in a genuinely wide container.

Platform sort headings are native keyboard controls

The shared platform sort heading declares column scope and contains a native non-submitting button. Enter, Space and pointer activation use the same existing three-step sort cycle; the whole heading remains a pointer target. Only the active column has aria-sort, and its visual direction marker is hidden from the accessible name. Non-sortable headings have no focusable control. Responsive labels are resolved once so direction updates cannot replace their DOM or erase their text. Column alignment, visibility, missing-value ordering and per-table sort persistence are unchanged.

PlatformSortableTableHead.test.tsx covers structure, stable responsive labels, focus, cycle/reset, active-only announcements and restored sorting. The real query/normalizer consuming TrueNAS completion tests retain the labelled control through filtering, an empty result and restoration, without inventing a finish for a running task. Native desktop keyboard and narrow touch behaviour require source-bound browser proof; component event dispatch alone is not that proof.

Visible action text remains in accessible names (24 October 2026)

Infrastructure section shortcuts include their visible Add action and group identity in the accessible name. Update chips include the visible singular or counted update label, the affected system and Agent Doctor destination, at both responsive layouts. Read-only inventories expose no add/update action, including when a caller retains an Agent Doctor callback. Patrol attention rows derive their name from their visible severity, title, summary, decision/resource context and observation time rather than a title-only override. Native buttons, pressed state, focus and touch selection remain unchanged. Registered InfrastructureSourceManager and PatrolAttentionWorkbench tests cover scoped actions, counts, read-only rendering and time-varying accessible content; the consuming WCAG scan and its rules remain enforced.

Agent first-report status

Machines, resource detail drawers and agent connection rows use the shared "Waiting for first report" label for monitoring's bounded pending startup state. Machines treats that informational state as unknown rather than an outage. The collapsed machine row shows the wording beside its name and suppresses cached metric bars until accepted telemetry arrives. Consumers must not infer the wait from a browser timer or hide a provider failure behind it. The expanded drawer presents the wait as information and does not synthesize a current chart reading from absent or cached metrics. Stored history remains available. The shared history query immediately withdraws cached live replies and current fallbacks when the owning source loses reporting authority. That transition starts a fresh read without waiting for the poll, while retaining actual stored series and rejecting a late live reply from the previous reporting state. Pending API polls retain their own wording.

Backups distinguishes completed points from inventory entries (24 October 2026)

The existing health strip names completed restore points and separately counts running and failed entries. The existing per-day chart measures only completed restore points in both count and volume modes, while the dated table and its filters keep incomplete entries available for inspection. Table and day-group counters name entries so they cannot imply completion. The incomplete-only chart state says that no completed dated points are visible, not that no backup activity occurred. The consuming Backups tests cover summary, chart, volume, date selection and route views together; existing table interaction primitives, access boundaries and routes are unchanged.

Backup posture failure and retry stay source-local (24 October 2026)

The same unavailable notice and Unknown projection apply when HTTP-success posture JSON has invalid rows, provider evidence or policy, duplicate subject IDs, or mixed policies across batches. Validation completes before publishing any batch. The existing retry remains disabled while pending and restores health only after an accepted complete read; independently readable restore points and open details stay usable. The parser, hook and consuming Backups tests pin this boundary; source-bound desktop/narrow pixels and keyboard/touch interactions establish the rendered result. No new control, route or shared query behaviour is introduced.

The Backups posture notice offers a keyboard-operable retry with the existing inventory-notice styling, wrapping action and disabled pending state. It refreshes only posture, not independently readable PVE/PBS inventories. Rows, health counts and expanded explanations withdraw failed retained assessments until a complete successful response arrives; independently observed restore points remain usable. Consuming table tests cover the notice and projection together. This changes no shared query cache, route, server policy or execution authority.

SSO editor feedback and exact group names (24 October 2026)

SSO save failures use the API client's human message/legacy error field, not the JSON envelope, code, timestamp or diagnostic details. Short plain-text validation errors remain useful; malformed, markup, control-bearing, oversized or unavailable guidance falls back to bounded recovery text. Raw response data and thrown network errors are not logged by this save path. A lost response must not claim that nothing was saved. Create and edit each make one write per explicit submission, including CSRF and rate-limit failures; the form, access restrictions and omitted saved secrets survive until deliberate correction or retry. The bounded, scrollable feedback area allows the complete ordinary initialisation/no-save guidance above the recovery controls on phones. SSOProviderSaveErrors.test.tsx covers both provider types and create/edit through the real state hook and API client. Presentation tests bind unsafe-input fallback; source-bound desktop/narrow browser proof uses captured producer envelopes and synthetic non-JSON gateway/validation responses. This is not installed sign-in, issuer compatibility, server persistence or original #3038/#61 recovery proof.

SAML metadata-preview failures use fixed guidance for recognised validation, fetch and parse categories, session/permission refusals and rate limits; unknown, legacy, malformed or unavailable replies use bounded connection/URL guidance. The editor never stringifies response bodies, error messages, diagnostic details or arbitrary callback values into this feedback or the preview logger. It names TLS trust on the Pulse server without suggesting a verification bypass. Each deliberate Preview makes one request, including CSRF/rate-limit refusals, and does not save provider configuration. URL/certificate drafts and access restrictions survive failure and retry; a successful parsed/XML preview still closes back to the editor and leaves saving explicit. Its existing Dialog return- focus capability targets the Preview button explicitly, even when disabling that button during fetch blurred it before the preview opened. No shared focus-trap implementation changes. Existing authentication expiry handling remains owned by the API client, not replaced by preview copy. The normal-suite SSOMetadataPreview.test.tsx exercises create/edit through the real state hook and API client, including failure -> explicit retry -> preview -> close/cancel, unsafe envelopes/HTML/network errors and disabled blank URLs. Presentation tests bind the category-only boundary without object coercion. Source-bound desktop/narrow browser proof uses existing public-shaped producer responses and adversarial safe fixtures through the production Settings editor. This is not a live IdP fetch, provider save, native sign-in or #3038 acceptance.

The provider editor scrolls its fields independently of its header and recovery footer. Save refusals, metadata-preview errors and clipboard outcomes stay in that active dialog, with live alert/status semantics; connection tests use the existing inline result. They do not rely on the intentionally inert background notification portal. Copy confirms success only after clipboard completion and otherwise offers manual copying. The full callback URL wraps at narrow widths. Save/Cancel stay reachable, and a pending save neither repeats nor dismisses. Only success-confirmation-only toast stacks lower below the modal backdrop while a blocking dialog is open. Any error, warning or info toast keeps the whole stack at its ordinary foreground layer, including when added alongside a retained success: existing delete, assignment and role dialogs still depend on those notifications while remaining open. Toasts are not unmounted or made interactive through the modal. Shared modal focus trapping, inertness and restoration remain unchanged. The SSO header and footer are non-scrolling flex siblings of the scrollable fields; the header needs no sticky positioning.

Allowed Groups are comma- or newline-separated exact IdP group names, preserving spaces through details → edit → no-op/mapping-only save for OIDC and SAML. This replaces the former whitespace split for that field only. Scopes, domains and emails keep their existing parsing; explicit empty restriction lists still clear them. Payload shape, secret omission and server-owned authorisation are unchanged. SSOProvidersPanel.test.tsx, ssoProvidersModel.test.ts, Toast.test.tsx, the consuming SSOProviderNotifications.test.tsx and the settings architecture suite cover this boundary, including an actual refused DELETE while its confirmation stays open and a successful save/reopen with the previous confirmation retained. Browser evidence uses the production panel, state hook, API client, Dialog, Toast and CSS with synthetic HTTP at desktop and 320px widths; it is not native IdP, encrypted persistence or restart proof.

Locked update pill names the command-access option by its shared label (21 October 2026)

For a host whose agent was installed without command access, the locked pill names the installer option through getUnifiedAgentCommandAccessOptionLabel ("Allow Pulse to run commands on this host") and the host to run the new install command on, instead of the retired "legacy combined command profile" wording and its advanced-options path.

Superseded 3 November 2026: the pill now says the host's agent "does not accept commands from Pulse" (an absent commandsEnabled reads as false, so it no longer claims how the agent was installed) and sends users to Settings → Infrastructure → Manage on that host → Remote commands, which names the option and opens the installer with it ticked. The pill no longer reads the shared label itself. containerUpdateBadgeModel.unavailable.test.ts and ContainerUpdateBadge.test.tsx pin the wording.

Container update control offers only what the server advertises (21 October 2026)

UpdateButton is live only when the server advertises the update capability and records no refusal for it. A recorded refusal, or no advertised capability (for example a host whose inventory is stale), renders the locked amber pill with the reason. When the refusal is a disconnected command agent and the host's agent facet reports no command access, the explanation names the host and the install option that grants access. The locked pill reads Update so it fits the Updates column at its narrowest; Local joins Pinned as a neutral uncheckable state rather than Check failed.

Aggregate disk bars preserve percentage-only evidence (14 October 2026)

The shared stacked-disk primitive keeps independently supplied aggregate percentages in percentage points, including values below 1%. Without a measured filesystem, it uses the aggregate metric admission in workloadDiskUsage.ts; capacity alone cannot draw a healthy zero. Percentage-only tooltips and labels withhold missing bytes rather than inventing empty capacity. Measured filesystem byte ratios, omitted-zero DiskInfo rows, unknown sentinels, per-filesystem thresholds and source-owned last-known/unavailable decoration remain separate.

Verification: the stacked-disk component/model suites retain filesystem and host controls; workloadDiskUsage.test.ts and the desktop/narrow workload-disk-metrics browser fixture cover aggregate low/zero/byte/unknown readings through the consuming Workloads row and drawer. A fixture is not a native measurement or stored History observation.

Pulse Mobile retirement guidance stays readable (14 October 2026)

The existing Pulse Mobile settings surface exposes its retirement guidance as a labelled note. An instance without the Relay capability uses the shared InlineNotice, not the sales gate's two-line body: the date and existing ntfy, Gotify and Pushover alternatives wrap in full on a phone. Existing entitled phones retain their configuration and pairing controls until 31 March 2027; neither note adds an offer, billing read, pairing action or new navigation.

RelaySettingsPanel.runtime.test.tsx preserves the full unlicensed message, absence of clamping/actions and existing entitlement controls. The first-session and paid-navigation browser tests target the actual labelled note rather than the desktop-only page subtitle. Desktop and narrow consuming proof must inspect the rendered copy and unchanged hidden-form/entitlement boundaries.

Independent PBS resource and inventory rendering (13 October 2026)

The existing PBS health/datastore/detail table remains mounted outside backup inventory loading/error gates. Explicit per-source notices and retries are independent too; both inventory failures retain their error state and never publish zero restore-point success. Known resource identity and readable capacity remain usable, while PBS backup counts stay Loading/Unavailable until a valid inventory arrives. ProxmoxBackupsTable.test.tsx verifies pending, single/both failure and ordinary recovery with a stable server-table owner. Desktop/narrow production-app proof must retain actual API resource identity, keyboard/touch detail access and truthful incomplete counts. No monitoring, backup execution, agent, credential or access authority is added.

Alert-first drawer composition retains usable evidence guidance (13 October 2026)

The guest's full OS caption moves to the visible, wrapping facts strip; it is not shortened or made tooltip-only when Summary moves into closed Details. Filesystem evidence and its existing reading-guidance link stay outside that disclosure, including unknown and last-known readings. The link keeps its 44px phone target, compact desktop target and keyboard outline.

The node's denied-update explanation and HelpIcon are owned by the drawer, outside the demoted Summary and the lifted warning rows. An ordinary same-node reading or relative-age update must not replace the open help dialog or its trigger. Update evidence wraps in Worth checking. Alert-first 24h cards, metric-owned alert colours, on-demand History, closed inventory Details and unchanged advice/actions remain. This is presentation/access continuity, not native QGA/backup or update-permission recovery.

NodeDrawerOverview.updateEvidence.test.tsx and GuestDrawerOverview.filesystems.test.tsx cover the connected merge outcomes; browser-tests/drawer-composition.cjs consumes the actual adapter, drawers, shared history query and Docs with synthetic observations at desktop/phone widths, including keyboard/touch guidance and same-node updates.

Control names retain their visible wording

Action names include their visible text in source order. Context may extend a short label, but must not replace it with a different action name. Infrastructure group shortcuts retain Add before their source label; update chips retain their visible count and update wording before the host context. Patrol attention rows use their native button contents so severity, title, summary and observation age remain part of the name as the displayed record changes. Scrollable command displays are native keyboard focus targets with visible focus. A sibling Copy button does not make the command's own scroll area keyboard reachable. Semantic muted text must meet normal-text contrast on the surface and hover backgrounds where controls and disclosures use it in either theme. Dark muted text uses slate-300 so slate-700 hover/consent panels remain readable; individual callers must not compensate for an inaccessible shared token. The authenticated browser accessibility proof covers both themes and an explicitly hovered Patrol row as well as expanded installer and Agent Doctor states.

Desktop and mobile utility links include their visible label and badge numbers contiguously, in source order, in the accessible name (WCAG 2.5.3). Mobile icons may appear above/beside their labels through CSS, but their numeric badges follow the label in the DOM with explicit word boundaries. Mixed Alert badge numbers remain in the name; the accessible description retains their critical/warning meanings without interrupting that visible text sequence. Patrol attention and Actions approval explanations stay in their names. Counts, routes, permissions, current-page state and stable link identity do not change. The maintained AppLayout/MobileNavBar/model tests and the integration WCAG scans cover these semantics, including zero, single and mixed severity counts. The integration graph keeps Playwright 1.56.1 while upgrading axe to 4.14.0; no WCAG tags, assertions, budgets or disclosure checks are removed.

History honours explicit selected-source staleness (13 October 2026)

The shared memory presentation reads the server's explicit agent.stale invalidation only when the selected observation is from agent. A dated current carrier then becomes visibly last known, with its original source, time and measured value, including zero. It does not supply current sorting, metric-filter pressure or guest/canonical History fallbacks. The visible source label explains that the supplying Pulse Agent stopped reporting. Missing or invalid provenance remains unknown; unavailable telemetry stays unavailable.

QGA and Proxmox observations remain independent of a stale linked Pulse Agent and of filesystem deferrals. No client TTL, source timestamp renewal, metadata rewrite, API schema, History point or guest operation is added. The source can restore current measured zero when it reports again. This is presentation safety, not native restart/thaw recovery or release availability.

staleAgentMemory.test.ts connects qualification to both History models, sorting and metric search, while GuestRow.test.tsx mounts same-ID source updates in bars and sparklines. The static production-component browser fixture guest-memory-agent-stale.cjs exercises agent-facet-only canonical deltas through the actual Workloads projection, both drawers and trusted keyboard/touch controls at desktop and phone widths.

Patrol setup panel uses existing form primitives

PatrolSetupPanel.tsx builds the provider choice from SelectionCardGroup (compact), the inputs from the shared form classes, results from InlineNotice, and actions from Button. It adds no shared primitive; at phone width the cards stack two per row and the action buttons keep 44px targets. The workspace mounts it through lazy and Suspense with no fallback, so the setup card shows the panel once its chunk loads. An unreachable provider keeps the existing setup card and ButtonLink, now pointing at Provider & Models.

Patrol header default-mode notice uses existing primitives

The Pro default notice in PatrolIntelligenceHeader.tsx is a role="status" block with two plain buttons that call the existing autonomy save. It adds no shared primitive, badge or banner variant, and it wraps on phone widths with 44px targets.

Canonical VM History respects native filesystem power state (12 October 2026)

The canonical resource drawer uses getWorkloadGuestDiskRead, the same VM filesystem admission as Workloads and Alerts. Proxmox runtimeStatus takes precedence over canonical status from a linked source; absent native state falls back to canonical status, with offline treated as stopped. A stopped VM exposes neither a current disk/usage fallback nor a last-known disk carrier, even with a prev- reason or refreshed Last seen. The stopped explanation is visible beside History. Running retained/unavailable reads keep their existing distinction; an accepted zero is still current.

Stored history remains inspectable, and independent CPU/memory/network metrics, LXC platform storage and non-Proxmox resources do not inherit VM filesystem rules. Power/read snapshot changes do not initiate a History refresh or a guest operation. This does not establish thaw, writes, restart recovery or release availability.

resourceDetailDrawerMetricsHistoryModel.branchcov0712.test.ts covers native and fallback state, platform scopes, zero, reason precedence and independent resource types. ResourceDetailDrawer.history.test.tsx connects reactive same-ID snapshots to the production drawer and History client. The production desktop Chromium/phone WebKit fixture browser-tests/resource-stopped-filesystem.cjs also preserves stored observations and trusted keyboard inspection.

Existing alert History distinguishes failed reads from empty results

The History tab renders fixed non-provider error copy in a role="alert" region and keeps its native Retry history button busy/disabled during the logical read. Loading is announced with role="status". Frequency and severity totals are hidden until a read is accepted; failed reads do not render the ordinary empty-history claim. Same-range saved rows and live alerts can remain under the warning, without presenting a different range's saved rows.

The connected History tab/state tests and parent-bound browser receipt cover deferred/synthetic failed reads, keyboard retry, accepted recovery and empty results, retained entries, range/context replacement and disposal. Desktop 1280px and phone 390px layouts fit without page-level horizontal overflow; the phone retry control preserves the shared 44px target and the desktop control keeps shared compact density. No new primitive, route, feature, permission, notification replay or backend recovery claim is introduced.

Existing webhook list controls expose pending acknowledgement

The webhook section shares pending admission across form saves, list toggles, bulk changes and deletion. Pending list writes announce a role="status" message and disable competing list controls and Add. An open draft also locks the saved-row controls; its own fields, Test and Save remain available until that draft is submitted. A fixed role="alert" message explains that earlier batch changes may already be saved when a later change is unconfirmed, rather than presenting an atomic rollback or automatically retrying it.

The existing section, state owners and API adapters are exercised in desktop and 390px browser fixtures with deferred/rejected synthetic replies, masked drafts, deliberate retry and unmount. List controls retain their narrow 44px touch targets and the section does not overflow horizontally. These checks establish source interaction and layout, not ordinary notification delivery. No new primitive, route, permission, destination or transport policy is added.

Existing alerts Save controls expose acknowledgement state

The settings Save button renders Saving… with aria-busy while its logical save is pending, and both Save and Discard are disabled until completion or failure. Newer edits remain possible and visibly unsaved; acknowledgement is not a claim that those edits were saved. Shared alert-configuration copy keeps the existing normal-save vocabulary and supplies a separate newer-unsaved message. Keyboard activation, tab re-entry and 320px/390px layouts are covered by the parent-bound browser fixture. alertConfigPresentation.test.ts and alertSettingsSave.acknowledgement.test.tsx exercise the copy and state owner. No new primitive, route or permission is introduced.

Threshold, schedule and destination editors appear only after saved policy and destination loading have completed for the current context. Pending reads render an announced loading message, not editable defaults. Failed policy reads render an alert and keyboard-operable Reload settings control; neither Save nor editable default fields is exposed. The copy and acknowledgement controls cover failed reads and successful reloads; the connected browser fixture also requires preserved pending-save drafts and narrow-layout fit. Overview and History navigation are not gated by this editor load state.

Existing Assistant control selectors describe planning

Settings and Chat have only Read-only and Ask first options. Legacy autonomous/suggest responses render Ask first without a permission warning or third execution mode; unknown values render Read-only. Settings compares normalised values before PUT so unrelated saves preserve a legacy preference. The two-option Chat menu keeps radio semantics, keyboard cycling/Escape/focus and saved-error recovery. Copy explains infrastructure plans in Actions and separate Patrol authority; it does not grant approval or execution. Existing scoped resource/alert/Patrol briefings and compact layouts remain intact. The panel has no Protected guests field: it never excluded a saved action plan, so it was removed, and a muted line points operators at the resource's Never auto-remediate lock instead (ai-never-auto-remediate-hint). The Chat control menu clamps to the composer and viewport on open and closes on viewport resize, so a wrapped toolbar does not clip either option. Mounted geometry checks and real 390px/320px browser bounds cover this observed gap.

Shared alert breach-date presentation

Existing attention and alert/history hover copy uses the validated metricStatus.lastBreachAt before legacy lastSeen, never occurrence start or poll time. Unknown/Go-zero dates stay undated. Shared relative formatting receives the presentation caller's clock for both stale-reading and breach ages. The existing compact Assistant attachment and approval boundary remain unchanged; only its model context gains the validated UTC date. Date controls in alert helpers and the production held-breach browser fixture bind this shared presentation boundary to the alerts/API contracts.

Alert severity fills are one primitive

getAlertSeverityFillClass in frontend-modern/src/utils/alertSeverityPresentation.ts is the one fill per alert severity, and getAlertSeverityDotClass composes it, so a legend dot and the chart segment it names cannot drift apart. Each fill holds at least 3:1 against the chart surface in both themes; yellow on the light surface is under 2:1, so light-theme warnings use amber-700.

Systemd backup-pause help separates shutdown from remote completion

The existing systemd precaution reads the stop result as well as the inactive service state. A newly stopped server requires a normal zero process exit; timeout, signal, non-zero or unknown results do not permit a planned backup. Previously inactive servers stay inactive, and failures must not be cleared to make the check pass. Even a successful shutdown does not cancel or prove completion of an already-issued guest-agent request. Existing guest/backup operations must finish normally, without new probes; unknown state, a disk dash, a cooldown or a waiting period is not completion evidence.

backupSafetyHelp.test.ts verifies the rendered bounded state command, failure and in-flight gates, prior-state restoration and shipped mirror. This changes existing help, not service configuration, native shutdown/QGA recovery or release availability. The independent post-backup thaw/write/liveness and only-prior-active restoration boundary remains unchanged.

Container backup-pause help preserves identity and prior state

The existing VM Disk Monitoring guide also covers a standalone Docker or plain Compose Pulse server, linked from the Docker guide and the systemd precaution. It identifies the original server container(s), discloses the monitoring/alert outage, requires suspended external restart controls and same-ID stopped-state readback before a planned backup. Stopped state never proves an earlier guest request completed. Incident recovery and controller-managed replacement are outside this manual sequence.

Restoration requires independent post-backup thaw, fresh successful workload writes on every covered filesystem and liveness. Only previously running, existing container IDs and previously active automatic jobs may be restored; unknown state, missing IDs, failed startup or unknown restart behaviour keep automation paused. No image, restart-policy, mounted data, guest operation or diagnostic permission changes. Bounded state templates omit environment/config secrets. dockerBackupSafetyHelp.test.ts exercises rendered commands, gates, cross-links and shipped mirrors; this is help validation, not a Docker daemon, installed guest recovery, native acceptance or release proof.

Narrow Workloads identity during guest locks

In narrow and phone layouts, the guest name and status occupy the first line; the unchanged backup evidence control moves below the name, retaining its 44 px touch target. The guest lock occupies a separate, wrapping line below that identity. It must not consume the name's horizontal space or spill into the adjacent metrics. The full lock remains visible; a long name keeps its full title and existing truncation. Wider layouts keep the inline lock. Disclosure, first-touch row activation and memory observation source/time are unchanged. This is presentation only: a backup lock or successful backup never proves guest recovery.

GuestRow.test.tsx checks layout structure, lock updates and existing action ownership. browser-tests/build-guest-row-lock-fixture.mjs compiles the real row, stylesheet and drawer before bounded static browser execution. Use node browser-tests/build-guest-row-lock-fixture.mjs --with-parent --parent <full-parent-SHA> to read the exact parent row from Git; no manually supplied control is accepted. browser-tests/guest-row-lock-identity.cjs measures the rendered name/lock at 320 and 390 px in bars and sparklines, and exercises keyboard/first touch through the existing drawer, checks long-name bounds, and rejects the current parent with the same static fixture. No dev compiler or old terminated preview is replayed. Native recovery remains separate.

Confirmed operations opt out of silent transport fallback

The shared API client accepts expectedOrgID and retry:false for confirmed Patrol manual-rule removal. It pins and rechecks the selected organisation at dispatch, respects cancellation after asynchronous CSRF preparation and makes no automatic auth/tenant/CSRF/rate-limit replay. Existing caller defaults are unchanged. Tests exercise these safeguards with real request options.

Patrol manual-rule confirmation keeps its title and safe Cancel/Remove controls outside a keyboard-focusable scrolling scope/reason region. Long multiline reasons cannot clip identifiers or the action footer on a narrow viewport. The region permits both keyboard and pointer review without changing confirmation state or sending a mutation.

All drawer History fallbacks must distinguish current observations from retained, unavailable or freshness-unknown guest readings, using original memory evidence and filesystem read reasons.

Workloads rows preserve memory observation provenance.

Workloads rows use the same selected-memory qualification as both drawers. Retained and freshness-unknown values keep a visible cue in bars, sparklines, the hover History lens and host-capacity mode; compact layouts shorten only the visible label, not the accessible original source/time explanation. Sparkline accessible values never call retained, unknown or unavailable memory current. An unavailable reading shows the cell's dash in bars mode (N/A only as a sparkline value) while recorded History stays intact; it has no visible label, because no number is left to qualify, and its reason stays in the cell title, the chart's native hover title and assistive text. An independent qualified memory reading is not paused by a disk deferral. Neither Last seen, backup completion nor running power state renews memory. Unannotated unrelated platforms and actual zero readings retain their meaning. Selected agent memory freshness is independent of VM/LXC classification. A stale agent retains its original observation as last known, while an independent current Proxmox observation remains current. Guest kind still determines the provider's cache interpretation.

A current reading from a cache-inclusive Proxmox source (status-mem, status-freemem, cluster-resources, derived-total-minus-used, the same sources and guest kinds the backend marks MayIncludeCache) keeps its numbers, colour and thresholds; the shared presentation adds mayIncludeCache and a cacheNote saying a high percentage alone does not show memory pressure and which fix can supply the guest's own figure. A VM row shows a cache cue, a question-mark icon beside the memory bar whose accessible name says May include cache (Cache? in compact layouts), only while the guest's own memory meets its warning or critical threshold, the figure that colours the bar in host-share mode too, which is when a cache-inclusive figure reads as pressure; below that the caveat is the cell title, the chart's native hover title and assistive text, not a label on every agentless VM. A container's cluster-resources reading is not cache-inclusive: Proxmox subtracts page cache from the cgroup usage (PVE::CGroup::get_memory_stat) before it reaches /cluster/resources, so memoryObservation.ts classifies that source by guest kind and neither row, tooltip nor drawer qualifies it. A QEMU guest, or a guest of unknown kind, keeps the caveat for it. Any other cache-inclusive container reading keeps the caveat in the bar tooltip and drawer instead of a label on every row. A retained or freshness-unknown reading keeps its own freshness label with the caveat in its explanation, an unavailable reading is never qualified, and the drawer states the note as visible text in About this reading. An unavailable explanation names only what is known: a powered-off guest reads Unavailable while the guest is powered off., and no message says Unknown source. Agent setup advice first states that setup, restarts and live guest-agent probes must be deferred during backups or a guest incident. This precaution stays with VM/container fallback guidance regardless of freshness or a reported backup lock; a fallback value or an ended lock does not establish that changing agents is safe. It changes no numbers, source selection, collection, thresholds or guest actions.

GuestRow.test.tsx, GuestDrawerOverview.memory-cache.test.tsx, MetricMiniSparkline.test.tsx and memoryObservation.test.ts pin these boundaries, with existing drawer, adapter and History suites. browser-tests/guest-row-memory-provenance.cjs checks real production rows and drawers, visible/accessible labels, pointer/touch/keyboard, light/dark and narrow/desktop layouts using synthetic observations only. Presentation acceptance does not establish native thaw, covered-filesystem writes, guest liveness, service restoration or reporter recovery.

Node drawer CPU temperature reads the backend reading

The node drawer's Thermals chart reading, its Summary temperature row and getNodeDrawerCurrentMetrics read Node.cpuTemperature, the backend's primary CPU reading, which the chart's stored line records under the same rule. getCpuTemperature (a frontend maximum over package, CPU maximum and cores) is removed from utils/temperature.ts; no drawer may rebuild a reading from temperature details. Proof: NodeDrawer.test.tsx (the chart's live fallback shows the backend value while the details hold a hotter core) and the nodeDrawerModel branch tests.

Drawer attention detail line and temperature alert floor — issue #2068

DrawerAttentionSection items accept an optional detail (a muted second line) and title (hover text). TemperatureGauge accepts alertSeverity and title, and getTemperatureTextClass treats an open alert's severity as a floor: a reading that has dipped under an open alert's trigger keeps the alert's tone, while the reading still wins when it is worse.

Alert card and incident panel format held alerts through one helper — issue #2068

AlertOverviewAlertCard and AlertResourceIncidentsPanel describe an open threshold alert through features/alerts/metricAlertPresentation.ts (summary, clear rule, alertLevel and clearLevel) instead of their own unit guesses or the breach message. useAlertHistoryState exposes its active-alert accessor so the incident panel matches its open occurrence to the live alert without a second store read, and the panel reads useRelativeTimeNow so a reading that stops updating turns stale on screen. From the sm breakpoint the card's left column (status icon and text) keeps a 16rem floor and the text breaks long words, so the action buttons wrap instead of squeezing the reading to its longest word beside them.

History rows read held alerts through the same helper — issue #2068

AlertHistoryTableAlertRow and AlertHistoryMobileList take their message copy from getAlertHistoryRowCopy in features/alerts/alertHistoryModel.ts, which wraps the same metricAlertPresentation.ts helper for an active row's liveAlert. It reads the useRelativeTimeNow clock only for rows with a live alert, so a reading that stops updating turns stale on screen while closed rows never re-render on the tick.

Canonical drawer History preserves guest read provenance

The shared resource drawer passes selected memory observation state/source/time and Proxmox VM disk-status reasons into the existing History deferred-reading path. Last-known values remain labelled evidence; unavailable numeric carriers never become current values. Unannotated Proxmox VM/LXC memory and unusable current timestamps are freshness-unknown. Non-Proxmox unannotated snapshots, independently current CPU/network/memory and stored observations stay unchanged. The memory presentation policy is shared with the dedicated guest drawer in memoryObservation.ts; it never borrows Last seen, a disk deferral or a snapshot refresh time. Unknown source/state text is not echoed.

resourceDetailDrawerMetricsHistoryModel.branchcov0712.test.ts pins both memory and filesystem paths, including invalid numbers/dates, zero, every deferral and unrelated platforms. ResourceDetailDrawer.history.test.tsx exercises the full reactive drawer with the real History renderer; GuestDrawer.test.tsx preserves the dedicated path. browser-tests/resource-history-provenance.cjs verifies these production components and clients with synthetic observations on desktop and touch WebKit. These are presentation checks, not independent thaw, native filesystem writes, safe service resumption or release/installed acceptance.

Observed CPU labels preserve small positives

Shared responsive CPU cells distinguish actual zero from small positive readings: positive values below 0.1% display <0.1%, other values below 10% retain one decimal without a trailing zero, and larger values keep whole-percent labels. Both compact text and animated bars use the same formatter. Explicit labels, non-CPU table formatting, unavailable/offline fallbacks, raw progress widths and threshold colouring retain their existing contracts.

Shared percentage History legends, inspection and accessible descriptions never round positive values below 0.1% to zero. Actual zero remains 0.0%; all other History units and precision stay unchanged. This changes displayed text only, not readings, coordinates, stored points, selection, requests or alert policy. Mounted responsive-cell and container row/drawer tests exercise these boundaries.

Guest drawers consume selected memory provenance

Existing guest drawers consume optional server-owned memory.observation and selected canonical/API metric annotations without renewing their original observation time from resource LastSeen, disk-read deferral or another platform facet. useWorkloads, useUnifiedResources and the legacy resource adapter preserve the selected annotation. A selected unannotated metric cannot borrow a conflicting raw platform annotation. Raw-memory fallback retains its own annotation only when its numeric reading is actually selected.

Proxmox VM/LXC Overview states current, last known, unavailable or freshness unknown with the original known UTC time and an allowlisted source label. Missing, unrecognised or unusable provenance is not current. Unavailable usage keeps known capacity but never invents zero used/free/swap measurements. History keeps retained/unknown numeric context separate from current legends and stored points, dated inspection and denied source access. Same-guest source changes update without resetting the drawer or refetching History. Independent valid PVE/Pulse Agent readings remain current even while QGA disk reads are deferred; unannotated unrelated platforms keep existing presentation.

Connected original-source/normalizer/adapter and mounted History controls plus browser-tests/guest-memory-provenance.cjs exercise these boundaries. They do not establish native thaw, covered-filesystem writes, liveness, reporter relief, publication or an installed containing release. No route, collection, numeric selection, retention, alert policy or diagnostic permission is added.

Pulse system-alert presentation uses the existing reserved ID/metadata markers, not the displayed name or a platform fallback. Overview and both History layouts keep service warnings, occurrence timelines and acknowledgement usable without a false resource link/policy action. History retains the system marker when reconstructing Assistant context; the explanation has no fabricated metric, guest target or targeted Patrol. Ordinary monitored resources named Pulse keep their original affordances. The shared scope helper, mounted alert controls and desktop/narrow-touch browser fixture cover this boundary. No navigation entry, page, API, alert policy or operator permission is introduced.

Guest-agent coverage is not inferred from metadata

The existing workload drawer distinguishes assigned Pulse Agent action ownership, observed agent version metadata and Proxmox-reported guest-read state. Neither a version string nor a parent-node execution target may be called connected. Reported available, deferred, expected-unreachable, not-running, disabled and unknown states remain visible even before a version or filesystem sample exists. A current operation lock/backup takes precedence over a cached available flag. QEMU read state remains separate from Pulse Agent metadata/assignment and does not appear for an unrelated vSphere VM or a system container from version alone. Unknown raw provider states/reasons are not printed in the new state explanation.

During a fixed guest-read deferral or uncertain completion, a visible, wrapping precaution remains outside the drawer tabs. It links the existing backup-safety guide, cautions against live diagnostics or restarting QGA during backup, and requires independent thaw and writes to covered filesystems. Running power state, an OK backup or reported agent availability establishes none of those. Fresh same-guest evidence withdraws the precaution without a new diagnostic or collector request. Existing filesystem provenance, metric values, History, action permissions and monitoring policy are unchanged. Memory qualification uses its own authoritative observation; disk-read state cannot qualify a different metric or renew its observation time.

Verification: GuestDrawerOverview.agent-state.test.tsx covers each current state with/without cached version, unknown private fields, action ownership, platform separation, lock/uncertainty/recovery and unchanged measurements. Mounted GuestDrawer.test.tsx covers the notice across Overview/History/Manage and the shipped-doc link without a guest check. browser-tests/guest-agent-coverage.cjs uses production row/drawer/router/docs and CSS, keyboard/touch, light/dark and 320px/390px/desktop, with synthetic responses. Source/render proof is not native thaw, safe writes, installed acceptance or a containing published release.

Backup inventory ownership is source-scoped and unambiguous

Both Backups views follow the storage source-association contract, including its source controls and qualification limits.

Diagnostics are live checks; downloads require a privacy review

The existing Settings Diagnostics panel explains before either run action that checks can make live API and guest-agent requests, must not run during backup, freeze/thaw or an unresponsive-host incident, and cannot prove ongoing collection has recovered. Both run controls reference that visible explanation. Export controls distinguish Full (private) from GitHub (review first) and reference the visible sharing warning. They download the displayed result locally without another diagnostics request or upload, and remain disabled during a refresh. The existing Troubleshooting viewer links to matching shipped guidance.

Sanitised exports redact the server's actual token usage IDs and agent names, memory snapshot identities and breakdown instances, nested guest IDs/names, mount paths and physical disk paths. Per-download aliases preserve actual token, agent and snapshot joins across differently ordered arrays. Counts, memory measurements, failure states and known fixed API failure categories remain useful; unrecognised upstream response bodies are redacted. Nested IPv4 prose receives the same IP redaction as top-level errors. The full/displayed result is unchanged. This is not an anonymity guarantee: arbitrary prose, future fields, credentials, secret URLs and other private information still require manual review.

Verification: diagnosticsExportPrivacy.test.ts exercises current server field names, reordered identity joins, nested privacy gaps, unchanged measurements and unchanged full data. Diagnostics presentation, hook, results and Settings tests cover the connected actions and warnings; shipped docs remain byte-synchronised. browser-tests/diagnostics-safety.cjs exercises production components, the real API reader and local downloads with synthetic responses on Chromium and touch WebKit, including loading/failure, keyboard, narrow layouts and the Docs link. These proofs do not run native guest checks or establish release availability.

PBS capacity warnings stay truthful and readable

The existing Backups server table distinguishes unavailable or unusable named datastore capacity from an explicitly measured empty store. Fixed Unavailable and Unknown notices replace unreliable percentages, bytes and deduplication figures without exposing a provider error body. Server reachability, independent backup inventory and drawer identity retain their own evidence and semantics. Measured zero and the existing high-capacity warning thresholds stay unchanged.

At compact widths, the shared 40% server-identity anchor stays in place; the Used column receives 25% so the warning remains readable without horizontal overflow. The redundant connection dot yields to its textual status. Datastore hover text does not change disclosure or row identity. The mounted capacity, drawer and full-page Backups tests plus browser-tests/pbs-capacity.cjs cover replacement, independent inventory, keyboard/touch and 320px/390px/desktop rendering. Synthetic presentation proof is not native datastore, backup, guest-thaw or release proof.

Backup age requires a usable completed timestamp

Existing workload indicators, age badges and drawer protection context distinguish an absent completed-backup timestamp from a present but unusable one. Finite, Date-representable positive timestamps no later than the current time may receive the configured freshness verdict. Future, malformed and out-of-range times are Unknown, cautionary rather than healthy or missing. The fixed explanation never renders the invalid raw value; future times direct the operator to check the Proxmox and browser clocks. An invalid current clock cannot confer freshness. Zero, empty and absent values retain the existing missing-backup interpretation.

Running activity remains separate: it neither validates the timestamp nor creates a completed backup. The badge, its tooltip and accessible name retain uncertainty while running; valid same-guest replacement restores the ordinary age verdict. backupTimeEvidence.contract.test.tsx, format.test.ts and workloadGuestPresentation.test.ts cover this admission and presentation. The real row/drawer browser fixture is browser-tests/backup-time-evidence.cjs. No backup execution, thaw, restored workload or installed acceptance is implied.

Filesystem capacity is not measured usage

The existing filesystem list and guest Overview require finite positive capacity and finite non-negative measured used bytes before displaying a percentage or utilization bar. Missing used bytes, invalid numeric fields, the negative unknown-usage sentinel and an overflowing ratio remain unavailable, not a healthy zero. Known capacity is still useful as ?/capacity; filesystem identity and type remain visible. An explicitly measured zero stays 0%, and omitted redundant percentage metadata does not discard valid bytes. Valid over-capacity readings retain their percentage and severity while the visual fill stays capped at 100%. No percentage or free-byte field is used to invent an absent used-byte reading.

filesystemUsage.evidence.test.tsx covers the model and both production consumers, including same-filesystem known → unavailable → measured-zero replacement. browser-tests/filesystem-usage-evidence.cjs covers these rendered boundaries on desktop Chromium and touch Chromium/WebKit. These synthetic observations are not proof of appliance collection, backup thaw, installed recovery or delivery.

Guest disk deferral and retained observations

The existing workload row, disk breakdown tooltip and guest Overview use one presentation for the fixed guest-read reason vocabulary. Operation locks, unverified lock clearance, an in-flight request, cooldown, incomplete response and capacity deferrals are not evidence that QGA is missing. Invalid VM identity remains distinct. A timeout describes uncertain completion and must not recommend restarting QGA during a backup or changing freeze settings.

prev- marks last-known disk values, not a fresh sample. Keep those values and their identity visible, with the explanation beside the disk evidence, including the breakdown tooltip and the phone-readable Overview. Show readable status even before the first filesystem sample; do not hide it behind an empty disk list or expose only a raw reason. A same-VM fresh observation withdraws the notice and updates values without remounting the row. Permission, disabled-agent and actual unavailable cases retain their own explanations.

When a linked Pulse agent's filesystems fill in for a VM with no Proxmox guest filesystems (disksFromAgent), the reading follows that agent instead of the Proxmox reason: current while it reports, last known once it is stale, and unavailable once the guest is stopped. getWorkloadGuestDiskRead in workloadGuestPresentation.ts owns that rule for the workload row's disk cell, the guest drawer's filesystem rows and its History, so the three cannot disagree about one guest; workloadGuestPresentation.test.ts pins it.

Verification: workloadGuestPresentation.test.ts, the mounted GuestRow, StackedDiskBar and GuestDrawerOverview.filesystems regressions, and browser-tests/guest-disk-deferral.cjs cover the fixed reasons, retained/no previous data and fresh resumption with production components and CSS on desktop Chromium and touch WebKit. Synthetic observations do not establish the native command-ID cause, guest thaw, fresh History collection or containing release.

The shipped row also keeps a filesystem provenance cue in every disk presentation: resting bars, selected sparklines, the pointer/focus history lens and the pre-sample empty state. prev- is labelled Last known (Prior in compact rows). An unavailable read shows the cell's dash with no visible label (a stopped VM or missing guest agent is a setup state, not news), except a broken read the user can fix: No permission or Agent error (Denied or Error in compact rows), in amber. Every reason stays in the cell title and the sparkline's native hover title. The full fixed-reason explanation is ordinary assistive text as well as the existing cell title and drawer content, not a live region repeated across a fleet. Retained sparkline values are announced as last known, never current; other metrics keep their existing current-value semantics. The inline sparkline plot yields width to the numeric value instead of reserving 40px and clipping that value in canonical phone columns. Clearing only a VM operation lock does not clear a reason still supplied for an in-flight read. Fresh linked Pulse-agent disk evidence with no reason remains unmarked, and non-VM rows do not inherit QGA guidance. The cue adds no action or tab stop; touch on it continues through the existing row disclosure to the full reason.

The opened guest Overview applies the same admission to every filesystem, not only the aggregate row or its status sentence. A valid prev- value is labelled Last known beside its mount identity and used/capacity bytes; retained values do not receive current utilization bars or threshold colours. Other VM read reasons and explicit disk telemetry unavailability show Usage unavailable, retaining identity and known capacity without promoting old used bytes. Missing or invalid retained usage is unavailable, not an empty healthy filesystem. The read explanation precedes the filesystem breakdown and remains visible before any sample. Clearing a lock alone cannot clear a still-supplied reason. A fresh same-guest read restores the normal accessible utilization bar; independent LXC and linked-agent readings with no VM reason retain their existing semantics.

Verification: GuestDrawerOverview.filesystem-provenance.test.tsx, the existing filesystem regressions and builder controls in DiskList.test.tsx, plus browser-tests/guest-filesystem-provenance.cjs using the full production guest drawer on desktop/light/dark and touch WebKit. These are synthetic presentation and input proofs, not native backup/thaw, covered-filesystem writes or resumption.

Verification: mounted GuestRow.test.tsx and MetricMiniSparkline.test.tsx regressions and browser-tests/guest-disk-provenance.cjs exercise retained and pre-sample states, display/lens changes, current-value semantics and same-row resumption. Native thaw, workload liveness and released acceptance remain independent obligations, not facts inferred from these synthetic observations.

The legacy Workloads DiskList is exercised only by tests/fixtures, not a shipped caller. Its empty-state title is not evidence of readable touch guidance; the shipped guest Overview supplies the readable pre-sample explanation.

Guest History applies the same provenance to its live legend fallback. A VM filesystem reason withdraws the current disk fallback; only a prev- reason with a valid retained value supplies a separately labelled last known value. Explicit telemetry unavailability and unknown/non-finite values stay unavailable, not zero. The readable fixed-reason notice remains beside the utilization group even when stored disk observations exist. Its wording identifies a live read, not the date or provenance of those stored points. Legend assistive descriptions reference that notice. CPU, memory and I/O evidence remain independent.

Deferred live evidence never supplies a chart point, timestamp or dated inspection value. Keyboard/pointer inspection continues to show only the selected stored observations, and denied History withdraws the whole chart including its deferred evidence. Fresh same-VM evidence clears the notice; lock clearance without a new read does not. GuestDrawer.test.tsx, GuestDrawerHistory.disk-provenance.test.tsx and GuestDrawerHistory.sparse.test.tsx cover the model, actual drawer plumbing, shared fallback precedence, retained zero, pre-sample status, dated inspection, access denial and resumption. browser-tests/guest-history-provenance.cjs uses the production full drawer and History API client at desktop and touch widths with synthetic responses. These checks do not establish native guest thaw, collection or release availability.

Backup age rules read through one resolver: resolveBackupThresholds in frontend-modern/src/utils/format.ts fills the fresh and stale defaults for both getBackupInfo and getWorkloadsGuestBackupAgeRule, so the copy that explains a shield (#2858) cannot disagree with the colour it explains. workloadGuestPresentation.test.ts pins the wording.

Completed guest backup evidence versus current activity

The guest Overview keeps the last completed backup age and its existing threshold-derived tone visible independently of current backup activity. Starting a backup must not hide absent or old completed evidence or turn it green. A separate cautionary activity row reports a running backup as not yet completed. Ending that activity without a new completed timestamp does not manufacture protection. A new completed timestamp updates the existing drawer without remounting it. Both values wrap at phone widths, rather than requiring hover to distinguish them.

Verification: the mounted GuestDrawer.test.tsx backup regressions, workloadGuestPresentation.test.ts and production full-drawer fixture browser-tests/guest-backup-protection.cjs cover VM/CT missing, stale and fresh completion evidence, concurrent activity, stopped activity with no completion, and later completed observations. They preserve independent disk-lock/timeout guidance and existing touch/keyboard collapse. Synthetic observations establish presentation, not native task success, archive restorability, QGA/thaw or release availability; those require their own containing native and delivery evidence.

Independent backup inventory read ownership

PVE files/snapshots/tasks and PBS inventory have independent, non-suspending read owners on the existing Backups page. A pending or failed source is not an empty successful inventory. Keep the other source's readable artifacts, filters and server health usable; name the missing source beside the evidence, mark restore-point totals as incomplete and retry only that source. If neither source can be read, the shared error action retries both. An initial pending read must not show “No backups yet” or a measured zero PBS count. Only a fulfilled empty PBS inventory establishes a zero count.

Pending retries retain independently readable evidence. A settled failed read withdraws that source's artifacts and counts until it recovers, without clearing the other source or inferring server-owned protection posture. A retry remains disabled and its previous warning visible while pending. Organisation switches withdraw prior inventory, replace/abort the read owners and refetch in the new API context; late responses cannot repopulate old evidence. Disposal aborts reads and removes subscriptions. This uses the existing query lifecycle, without new inventory polling or a remount cache.

Successful HTTP responses still need a readable backup envelope, collections and consumed row fields. Invalid JSON, missing data and malformed rows produce a source-local unavailable state with fixed format guidance, never raw response content, a fabricated empty inventory or a page-level rendering failure that hides another source. No invalid rows are silently dropped, no completion flags are coerced, and only a validated response clears the warning. Explicitly nil PVE collections remain compatible with the existing Go encoding. proxmoxBackupInventory.test.ts, mounted backup-table controls and browser-tests/backup-inventory-format.cjs verify these boundaries and isolated keyboard/touch recovery, not native backup or released availability.

Verification: mounted ProxmoxBackupsTable.test.tsx covers source pending, partial/both-source failure, isolated retry, fulfilled empty counts, 401/403, organisation replacement, late response and disposal. Production component, router, API adapter and stylesheet fixture browser-tests/backup-inventory.cjs checks desktop Chromium and touch WebKit recovery, warning readability, keyboard or touch retry, source isolation and actual new organisation headers. Synthetic HTTP results establish frontend truthfulness/lifecycle, not native PBS/PVE permission, archive restorability, guest thaw or released availability.

Composed operator tables and conservative rollback consent

Phone storage tables use the canonical weighted-width helper and matching header/cell gutters. Compact labels shorten disk verdicts and capacity values, not their underlying health or measured units; full reasons stay in the real disk header and tooltips. Structured ZFS scan activity supplies a compact badge only during reported rebuild activity; its complete provider summary remains available. Stable disk History catalog/organisation/access ownership is unchanged.

The physical-disk verdict words and their phone forms are Needs Attention (Attention), Running Hot (Hot) and Replace Now (Replace), all from getPhysicalDiskHealthCompactLabel in diskPresentation.ts. Running Hot is red and starts at the disk's alert trigger, from the same thresholds object that colours its Temp cell (getPhysicalDiskTemperatureThresholds). Callers pass the override keys of the machine that reports the disk, so that machine's Disk Temp override sets those thresholds.

Controller expansions carry the supplied kind-specific targets, absolute times, duration and cluster/namespace when the phone row omits those columns. Proxmox coverage distinguishes independent backups from guest-local snapshots; By date retains archive identifiers in hover/search while showing the format as detail. Phone identity weights leave every configured metric column visible.

UpdateHistory consent names both the requested target and the observed running version, separately from the version on the selected history entry. The API does not expose backup restore scope: do not infer it from a version cutoff, notes or a retained path. Explain installation-only versus legacy recovery and stopped, consistent full-state recovery without promising later settings will be reverted or preserved. Opening/cancelling never starts a write, rejection retains consent, and a pending request cannot be submitted again or cancelled.

Verification: the mounted UpdateHistory regression tests, existing kind-specific controller/storage/backup/workload tests and production-CSS browser fixture frontend-modern/browser-tests/operator-tables-consent.cjs exercise the composed source at desktop, intermediate and phone widths. Synthetic HTTP/snapshots do not establish native collectors, tenancy, installation recovery or release delivery.

Organisation-scoped single-metric History

Fetched HistoryChart observations belong to the active organisation as well as the resource type/ID, metric and range. org_switched replaces the request owner even when those chart props are unchanged: withdraw old samples, provenance, errors and touch/keyboard/group inspection, abort superseded reads and replace the polling timer before immediately requesting the same selection in the new API context. Late success or failure from an earlier organisation cannot update or settle that request. Disposal removes both timer and org subscription.

Same-organisation transport/5xx failures still retain their own successful observations with explicit refresh warnings. A failed first read in a new org must not fall back to the previous org. Locked or missing targets do not fetch; explicit supplied samples remain caller-owned, not silently converted to fetched history. Matching same-org supplied-sample refreshes retain inspection. Runtime ownership and keyboard/touch rendering regressions plus the production browser-tests/history-org.cjs HTTP fixture verify these boundaries with the real API org header and the existing org-safe batch reader as a control. This is client presentation/isolation proof, not native tenancy, collector or release acceptance.

Stable physical-disk History composition

The disk drawer Overview (DiskDrawerOverview, also shown for a guest's own disks) composes the shared drawer history from stable disk-family catalog entries, independently of the latest optional temperature/SMART fields. Matching snapshots must not recreate chart DOM, reset inspection or restart reads. Stored series retain their own sample provenance; current collection unavailability is still reported beside the SMART facts on Overview (there is no History tab), and empty history is not a manufactured zero. The shared target/range owner continues to cancel old selections and reject late responses. DiskDrawer.history-collection.test.tsx and the production browser-tests/disk-history-collection.cjs fixture verify these boundaries with synthetic data; neither establishes native NAS health or installed acceptance.

Storage pool Configuration key/value rows keep capacity absence distinct from measured zero: unavailable bytes and percentages render as n/a, while valid independent observations survive a missing total. Live partial snapshots must not invent an empty pool or available space. The existing shared row layout and responsive labels are unchanged; the production drawer browser fixture frontend-modern/browser-tests/pool-capacity.cjs verifies desktop and phone transitions and overflow alongside storage capacity regression tests.

Shipped documentation table headers

After sanitization, the documentation renderer assigns trusted scope=col to native thead th cells without expanding the attribute allowlist. Document roles, scope and classes remain untrusted, and body cells are not promoted. Table-local horizontal scrolling is preserved. Renderer regression tests cover trusted column scope and rejected author-supplied attributes. The production Docs/router/styles fixture checks all eight current plan-table columnheader roles in desktop Chromium and phone WebKit, including local scrolling and page-width containment. This establishes browser roles, not screen-reader speech or release availability.

Shipped documentation fragment navigation

The shared documentation renderer assigns GitHub-compatible, document-local heading IDs from sanitized text, retaining explicit anchors and avoiding duplicate IDs. The documentation viewer follows fragments after asynchronous content rendering as well as in-page navigation. Fragment targets receive keyboard focus without entering the normal tab order. Missing or malformed fragments do not throw or move focus. The renderer and fragment helper are covered by frontend-modern/src/features/docs/__tests__/docMarkdown.test.ts, with direct-link, reload and keyboard navigation verified in the live viewer.

Desktop table rows stay single-line

Platform, Machines, alert threshold and Settings > Infrastructure desktop rows are one line of at most 33px, the height of the platform tables. Controls taller than the 24px line overhang the row with negative margins instead of setting its height: the Machines actions button (sm:-my-0.5, 37px rows before), the Settings > Infrastructure Manage actions, and the alert threshold toggles and inputs. Settings > Infrastructure body cells use the platform tables' 4px vertical padding (39px rows before), and its system line keeps the name, collection badge and member toggle on one line, truncating the name, where a wrapped badge had made a 63px row. Long text ends in an ellipsis with its full value in the title. Alert History rows are one line from a 50rem card once its icon row actions and compact Status give Message and Resource the width; below 50rem Message keeps two lines.

Visible phone row disclosure keeps rows single-line

SummaryRowActionButton keeps a 36 px disclosure target on phones, where the Proxmox nodes, Docker hosts and vSphere hosts tables show it because the row tap scopes the table below. Negative margins let that target overhang the 14 px chevron, so the button no longer sets the row height: those rows had been 37 px beside the single-line 32 px rows of every other platform table. From the sm breakpoint the margins reset and the 24 px button is unchanged. SummaryRowActionButton.test.tsx pins the classes.

Patrol mode route anchor opens its disclosure

Patrol mode entry points in Settings and the licence panel link to /patrol#patrol-control, while the Patrol header keeps the Patrol mode selector in a collapsed Mode and automation disclosure. Arriving on that anchor, with or without the starter query, or on the #operations-loop compatibility anchor opens the disclosure and scrolls it into view. The header acts again only when the hash or the disclosure element changes, so a disclosure the user collapses stays collapsed and a plain /patrol visit leaves it collapsed. Plan-locked installs render no disclosure. frontend-modern/src/features/patrol/__tests__/PatrolIntelligenceHeader.render.test.tsx pins this on a rendered header.

Disk mount scrolling

DisksCard keeps every supplied mount in its parent's scrolling flow. It must not cap the mount list or create a nested scroll target whose only overflow cue is a platform scrollbar. Large lists intentionally increase card height; aggregate usage, individual mount data and empty-state behaviour remain unchanged. This boundary is local to DisksCard, not a global scrollbar styling requirement. SharedPrimitives.guardrails.test.ts protects this composition; component tests preserve mount counts and totals. The disk-mounts qualification fixture checks short/long lists, themes and keyboard reachability with production CSS.

Retained disk usage

DisksCard takes an optional lastKnownReason for usage figures that are retained rather than current. The machine drawer passes the host-agent stopped-reporting reason when agent.stale is set, the same signal that marks its Thermals rows "(last known)". The card keeps every figure, because a disk that filled before the machine went quiet is evidence. The total and each mount read "Last known N%" with used and total bytes, in muted text titled "Last known reading, not current: ". Threshold text colour, the aggregate StackedDiskBar and the per-mount bars are dropped, as the guest drawer does for retained filesystem rows. Mount rows may wrap, with the label capped at the row width, so on a phone a long mountpoint takes its own line instead of squeezing the figures out of the card. DisksCard.test.tsx covers both readings and SharedPrimitives.guardrails.test.ts pins the branch and the drawer wiring.

Retained RAID state

RaidCard takes the same optional lastKnownReason, and the machine drawer passes it on the same signal as the Disks card. The card keeps every array and member state, because an array that degraded before the machine went quiet is evidence. Each state reads " (last known)" in muted text beside a muted dot, titled "Last known reading, not current: ". A rebuild figure is not progress once the agent is silent, so where the card shows one (above 0 and below 100%) it reads "Rebuild was at N%" without the rebuild speed. Member badges drop their status colour, so a member that was not healthy names its state in the badge ("/dev/sdb2 · faulty"). The Machines table's RAID column applies the same rule from agent.stale to what it shows: a muted summary and a tooltip headed "RAID Arrays (last known)" without the rebuild bar or speed. The array header row wraps, so on a phone the longer state takes its own line instead of truncating the array name. The card's empty guard is tracked, so a drawer opened before the first RAID report shows the card once arrays arrive. raidPresentation.ts owns member health: mdadm's healthy "active sync" members read healthy, where they used to take the amber warning badge. RaidCard.test.tsx, raidPresentation.test.ts and AgentsMachinesTable.test.tsx cover both readings and the transitions between them, and SharedPrimitives.guardrails.test.ts pins the branch and both wirings.

Ollama credential editing

The provider panel exposes the existing Basic Auth configuration. Saved passwords are represented by presence text, never a placeholder secret or input value. Blank password input preserves the saved value; explicit clearing takes precedence over a draft replacement. Successful saves discard the password draft. Username clearing is independent. Password bytes are not trimmed. Tests exercise preservation, replacement and clearing through the Settings save action. Connection testing uses saved settings; the panel tells users to save before testing and use HTTPS remotely.

Assistant owns composer registration and focus on every open, rather than only on component mount. Closing clears the registered input so later keyboard commands cannot target a detached composer. A handoff must leave Escape and keyboard input in Assistant, not the underlying alert search.

Assistant mention suggestions preserve each resource.id. Provider scope and current node identify same-label guests where policy permits them, and search can use those visible fields. Selected chips retain independent canonical IDs, allow removal without discarding prompt text, and show current owned labels after inventory refresh. Governed labels never expose node or site metadata through either suggestions or chips. Prompt history keeps the same identity, including older bounded payloads interpreted by the runtime.

Mobile incident drawers transfer their exact context to Assistant and close through the shared explicit handoff callback. Keeping the source drawer above Assistant, or dropping its occurrence identity to make navigation work, fails the linked investigation journey. Both timeline and resource handoffs require mounted regression and final-build narrow browser proof.

Shared incident evidence disclosure

Alerts timeline and resource-history events share IncidentTimelineEventCard. Its native details/summary disclosure preserves keyboard activation and keeps forensic provenance out of the default event summary. The same timestamp formatter rejects missing or invalid evidence times. Timeline and resource history compose persistent failure copy with their existing retry controls, independently of transient notification toasts. The affected interaction and viewport qualification is recorded in docs/qualification/PATROL_ASSISTANT_CUSTOMER_JOURNEY.md.

Canonical Patrol and Assistant continuation, 2026-09-07

Patrol's Assistant context preserves unknown destructive risk and distinguishes canonical action state from legacy approval state. Transcript scrolling is owned by the shared Assistant message container, including streaming and Latest, so it cannot move outer document ancestors. Browser proof must inspect the header, composer, nested evidence and scroll position after streaming and viewport resize. The current scoped matrix is recorded in docs/qualification/PATROL_ASSISTANT_CUSTOMER_JOURNEY.md.

Recovery feedback composes shared controls without a timer

The alerts-owned AlertQueueActionFeedback composes Card and Button rather than altering global toast lifetimes. Its polite atomic status region is mounted before failure text arrives; the warning card uses semantic foreground and wraps text and the explicit “Clear recovery message” control at narrow widths. The enclosing labelled region remains mounted and receives focus before clear removes the button, avoiding focus loss to the document body. Updating feedback does not steal focus. This is view-local feedback, not durable delivery history.

The live region stays independent of delivery-health conditional rendering: a later healthy observation can remove the health warning without removing failed-action information. AlertDeliveryHealthCard.test.tsx verifies status and focus composition. scripts/check-recovery-feedback.mjs verifies both real feature views, keyboard activation/clear and text/control containment at 1440, 900 and 390px, with scripted responses and genuine toast expiry. It does not establish screen-reader announcement quality or installed notification delivery.

The shared action evidence disclosure preserves the named observer independently of the executor. Its observation timestamp uses the neutral label Observed, followed by the separate Pulse receipt time. Independent Proxmox API evidence must not be labelled as an agent observation. Browser qualification expands this disclosure in completed action reviews at desktop and narrow widths.

Disk I/O presentation preserves each observed direction independently. Shared formatting renders a missing rate as a dash and measured idle as numeric zero. Partial observations cannot form a complete throughput total for sorting or comparison. Machines column preferences must preserve an explicit user choice across the first reload, including default-hidden migrations. Final-source browser proof covers Docker host details, Machines column selection and tooltip focus/dismissal at desktop, intermediate and narrow widths.

Overview delivery diagnoses use latest-started refresh ownership. Older bulk responses cannot overwrite newer card notification status, and an empty active alert set invalidates outstanding reads. Disposal also prevents updates. Failed refreshes retain the existing snapshot; this ordering repair does not add a freshness indicator or establish recipient receipt. Verify response overlap in OverviewTab.deliverystatus.test.tsx, empty-set invalidation in useAlertOverviewState.test.tsx, and rendered ordering at three widths using scripts/check-alert-diagnosis-ordering.mjs.

The Destinations delivery-log state primitive assigns a generation to each refresh and rejects stale completions before updating rows, unavailable state or loading state. Held-event reads share that generation without blocking the attempt-log spinner. Cleanup prevents abandoned requests from updating state and makes subsequent calls through the disposed loader inert; it does not cancel transport requests. Latest-request failure remains unavailable rather than being concealed by an older successful response. Verification combines the hook's ordinary race/disposal tests, registered mount/Retry integration, and scripts/check-delivery-log-ordering.mjs Chromium content assertions at desktop and narrow widths. This is component-level presentation proof with scripted APIs, not full-tab or installed qualification.

The shared delivery-health card wraps action groups according to available space, retaining readable explanation width when Review, Retry, Dismiss and Refresh appear together. Its heading uses the opaque semantic foreground, not the translucent palette shades reserved for status backgrounds. Verify light/dark layouts at desktop, intermediate and narrow widths, including unavailable health, pending refresh and recovery.

The alerts overview offers the existing delivery-status refresh control when health is unavailable, including after a successful retained-queue action whose follow-up health read fails. The warning remains until a verified healthy read; a successful queue action alone is not evidence of delivery health. Normal degraded summary presentation continues to omit refresh.

Proxmox backup presentation treats every manifestless PBS artifact as non-recoverable. It renders the artifact as Running when current writer visibility is absent or a matching writer is active, and as danger-tone Failed when a complete current-task observation finds no live writer. Running and failed/incomplete artifacts remain inspectable but cannot become a workload's latest recoverable point; search includes running, failed, and incomplete state vocabulary. The compact recovery table reserves enough of its fixed phone-width layout for the complete state badge instead of clipping that recovery answer at the horizontal scroll edge.

Own reusable frontend primitives and canonical page-shell patterns so feature work extends shared components instead of creating new local variants. Feature-owned warning cards, including notification delivery health, compose the shared Button variants for retry, dismiss, refresh, loading, and disabled states. Feature code owns the action copy and confirmation consequences, but must not recreate local button chrome for those controls. Feature panels embedded in a shared Dialog or drawer must not duplicate the overlay's accessible heading. A reusable panel may suppress its standalone title when the owning overlay supplies the canonical title, while preserving that title in inline and desktop contexts; the overlay remains responsible for one visible heading, its accessible label, dismissal, and focus return. The shared Dialog component requires exactly one accessible-name strategy at its component boundary: consumers provide either ariaLabelledBy for a visible heading or ariaLabel when no visible label is available. Unnamed dialogs and consumers that provide both strategies must fail the frontend type boundary. The alert schedule's initial-delivery selector composes SettingsPanel and FormSelect, uses the shared alert-configuration presentation vocabulary, and exposes the same email, webhook, Apprise, and all-destination labels used by escalation. It must not introduce a page-local select shell or a second destination-label map. The centralized Findings surface loads active and historical Patrol findings, keeps operator notes editable (including clearing a note with an empty value), and exposes Reopen finding only for dismissed rows. Reopening uses the finding-backed suppression removal API, refreshes both unified and Patrol history, and must not discard the saved note.

Platform-owned workload controls extend the shared WorkloadsFilter view options rather than creating page-local toolbar shells. Persistent presentation choices compose the shared ViewOptionsDisclosure instead of occupying the primary filter rail: layout, metric style, row hover, memory basis, inventory totals, and columns remain discoverable behind one View trigger. The history range stays inline in both metric modes because bars now expose an intent-driven row history lens and Trends keeps the same charts persistent. In Bars the range renders only at desktop widths while Row hover is set to History, the mode that drives that lens; Details hover and the phone (isMobile) layout hide it. The inline range carries the visible History label that WorkloadsFilter passes to MetricHistoryRangeSegmentedControl. Controls inside the View disclosure must expand in place rather than opening nested absolute panels that can clip or create competing overlay stacks. The Proxmox page owns and persists the Guest / Host memory basis; the workload state, table, panel, and row contracts carry the selected basis and resolved parent-node data to the canonical memory bar, and the memory column header must expose the non-default Host basis after the control closes. The shared ColumnPicker may also expose Reset widths when an owning table has active manual column sizing. That action is separate from restoring column visibility defaults: feature state owns the width reset callback and active flag, while the shared picker owns the discoverable menu placement and button presentation. Tables without manual sizing omit both properties and retain the existing picker unchanged.

The default Workloads metric presentation keeps compact progress bars at rest. A fine-pointer preview or keyboard focus on one guest row replaces CPU, memory, disk, Net I/O, and Disk I/O together with the existing MetricMiniSparkline presentation without changing row height; touch pointer entry does not trigger this transient lens, and the persistent Trends View choice remains the touch-accessible fallback. Both I/O cells render through one rate renderer in the lens and in Trends: the chart keeps a compact current pair beside it (MetricMiniSparklineRatePair, the at-rest ↓/↑ and R/W glyphs coloured like the series they label) while the full rates stay in the chart's accessible label, and a stopped guest or unavailable I/O keeps its dash. The active chart owns its local tooltip while its normalized cursor position is shared across sibling charts in that guest row, so every guide represents the same relative point in the selected history range. Leaving the row clears the cursor and restores the bars and I/O readouts together. The lens mounts with a short reduced-motion-safe fade and must not leave both bar and chart semantics in the accessibility tree simultaneously. Bar mode resolves history per guest through each guest's canonical metrics target and the selected compact range, warming only a bounded window of mounted and adjacent rows with the active guest first; it must not start an estate-wide chart request merely because the range changes. The request key is stable across equivalent live guest snapshots, and selecting another range aborts superseded browser work. Leaving a row does not cancel its read, which settles into the bounded row cache. Persistent Trends may retain the shared estate reader, but range changes must clear prior-range data unless an exact-key cache entry exists.

Feature-owned scope controls that use the shared filter rail must keep their state in the owning route and use stable, domain-authored option identities. The Proxmox Backups Backup location control composes the shared filter catalog, reads PBS instance plus datastore identity from the recovery model, and persists unchanged between the By date and Coverage views. Clearing the shared filter rail must remove that route value along with the other active facets; the feature must not replace the shared rail with a page-local select. The TrueNAS Storage Storage type scope follows the same boundary with stable volumes and disks route values. It composes the shared filter bar, exposes the current option through pressed-state semantics, preserves the scope in the URL across reload, and removes the query value for the default all state. On narrow screens the physical-disk scope may reprioritize its canonical table columns to endurance and temperature while retaining the shared table overflow and touch-target behavior; it must not introduce a second mobile-only filter or table shell. Its Health column stays desktop-only in every scope: the status dot carries a disk's condition on a phone, the row drawer carries the reason, and the old pill clipped past the phone container edge.

Large-estate platform pages must keep one canonical inventory snapshot for the initial read and explicit refresh path. The Proxmox and VMware vSphere overviews own their source-scoped unified-resource requests and pass those snapshots into the shared workloads state; the workloads adapter may map the snapshot into the legacy guest boundary, but must not issue a second workload inventory request or create a second infrastructure poll. Refreshing an overview must invalidate that owner snapshot and update both the host/node and guest regions from the same result, so a large estate cannot render contradictory counts, flash a false empty workload state, or pay duplicate transport and reconciliation costs. Provider workspaces must also constrain their canonical query at the source boundary instead of downloading same-type rows from unrelated platforms and discarding them in the page model. Docker and Kubernetes use their provider source directly, including merged agent rows that carry that source. Proxmox keeps separate route-family queries, but only the active route may enable its query: inactive desktop and phone tabs must not create a background inventory burst. The Backups route composes the existing Overview guest snapshot with a PBS-only addition rather than issuing a second guest-estate request. The overview's bounded structural summary remains ahead of its long virtualized inventory at desktop and narrow widths. In particular, Proxmox must show up to six phone rows or twelve larger-layout rows before the guest list instead of visually moving the nodes after the guest list's full virtual scroll extent; estates at or below the applicable threshold render in full without a continuation control, while revealing a larger node estate remains an explicit table-preview action.

App-shell navigation tab lists rendered through reference-keyed <For> consumers keep stable item identity across websocket state frames. AppLayout derives its primary and utility tab arrays through the shared frontend-modern/src/components/shared/stableNavTabs.ts reuse helper, which returns previous tab object references (and the previous array identity) when a rebuilt list is structurally unchanged, so an alerts-bearing state frame with unchanged badge content cannot recreate nav button DOM and drop an in-flight tap. New nav or tab-strip consumers that rebuild their item arrays from live store reads must route through the same helper instead of <For>-ing over freshly constructed objects.

Estate-sized table and card rendering routes through the shared PlatformWindowedRows, PlatformWindowedList, and usePlatformWindowedItems primitives. They preserve the complete filtered and sorted result plus native scroll extent while bounding mounted DOM to 140 items on wider layouts and 36 on phones unless a feature declares a smaller budget. Wheel projection may prewarm a directional keyed-row runway before native scrolling exposes it. Touch scrolling must remain compositor-native: windowed renderers must not attach touch listeners or replace keyed rows before the browser moves the page, and must update their runway only from the passive native scroll event. Spacer geometry is structural only: no feature may present it as loading, pagination, or an intentionally blank data region. Settings resource pickers, Availability target lists, Actions, alerts, and every provider-native platform table share this contract. For table rows with unique logical ids, or an explicit unique key extractor, PlatformWindowedRows owns a stable wrapper and independently reconciled store per logical row. A live snapshot may reorder those wrappers without remounting row-local input or drawer state, and must never reconcile one row's nested value through another row or duplicate rows after sorting. Missing or duplicate keys retain the reference-keyed fallback.

The window's item-height estimate is measured from representative content, not the leading sibling alone. Grouped surfaces render a short group header before their first content row, and sampling only that header collapses the estimate so the mounted window advances far faster than the real scroll position and drops a group's rows mid-scroll. The controller samples several leading siblings and keeps the tallest, so uniform tables still measure their real row height while mixed group/content lists keep a content-scale estimate. Native table rows additionally measure each mounted item's summary and any InlineDetailTableRow marked data-inline-detail-row as one logical item. Keyed measurements own prefix offsets, native spacer extent and pixel-to-item lookup, so scrolling within a tall drawer cannot advance past and unmount its owner. Offscreen items retain their last measured size until remeasured. Resize and direct-child mutation observers update those measurements outside the scroll/wheel path. The mounted-item budget and native touch path remain unchanged. List surfaces retain representative height estimation.

Shared workload, node, Docker-host, and resource-drawer history presentation must scope retained observations to the exact resource type, resource ID and range. An uncached target or range change clears the former points while its read is pending, including when a PBS host link is withdrawn. A failed replacement must never cache former-host points under the new target. Matching cached reads and same-source background polls retain their chart without a loading flash. Superseded, locked, unavailable and unmounted requests propagate the query's abort signal to the Charts API; late results cannot replace current observations. GuestDrawerHistory.source-isolation.test.tsx exercises the real renderer, cache readback and cancellation, not a mocked chart. The mock-backed PBS browser runner verifies delayed range and withdrawn-target reads at desktop and phone widths; neither proof establishes installed collection or #1723 relief.

Failed same-source history refreshes keep valid previously loaded observations visible with an explicit warning, rather than hiding the entire chart. That retention applies to transient transport/server failures, never a final HTTP 401 or 403. Both the batch drawer and shared canvas reader withdraw stored samples and inspection on access failure; denied history is not an empty, collecting, or stale-but-still-readable result. Their guidance is bounded by numeric status, not transport bodies or guessed diagnostic text. The shared canvas announces its error, and the drawer keeps its scoped, focusable retry. Only a successful replacement read restores observations.

The retained-value query helper also withdraws its active value on a final 401/403 and clears its shared remount cache. Other resource/range entries were read under the same access context, so they cannot be resurrected after that context is denied. Advancing the cache generation rejects cache writes from already-running pre-denial reads. A superseded denied completion cannot clear the current target or its cache. This does not broadcast a session reset or claim revocation of independently mounted consumers; their current response still owns their presentation. Transient failures keep the existing retained value, and a newly successful read may populate the cache normally. Verification uses runtime assertions in createNonSuspendingQuery.test.tsx, GuestDrawerHistory.refresh.test.tsx, useHistoryChartState.test.tsx, HistoryChart.test.tsx, and browser-tests/history-access.cjs (both production readers/client, real same-origin 403/503 responses, cache/range/remount and fresh-read recovery on desktop Chromium and touch WebKit). These controls do not establish appliance authorisation, collection or released acceptance. Initial or uncached replacement failures show unavailable history, never borrowed points or a collecting claim. The existing target/range owns every manual refresh; locked or absent targets expose no refresh control. The control remains mounted and focusable through retry and recovery, rejects activation while busy, and updates a pre-mounted polite status region without exposing transport diagnostics. The latest query read settles loading even when background polling supersedes a foreground refresh; late superseded results remain inert. Verification: GuestDrawerHistory.refresh.test.tsx, createNonSuspendingQuery.test.tsx, and browser-tests/pbs-history-refresh.cjs (direct production History renderer, scripted failures, keyboard retry, overlap, target withdrawal, phone/desktop and light/dark themes). These proofs establish presentation/recovery, not installed collection or delivery.

Shared workload, node, Docker-host, and resource-drawer history presentation keeps current readings separate from stored samples. A current metric may populate the legend only with the explicit current label, never as a stored reading, timestamp or chart geometry. A successful empty read says No stored history in this range; it does not imply that collection is running or that changing provider configuration will produce a sample. Zero remains a valid reported reading, while an absent metric remains unavailable. When a panel has stored observations, its default legend uses the newest stored timestamp across that panel and shows only values actually observed at that time. It labels the time Latest stored; missing sibling series remain dashes, with a visible explanation, rather than borrowing an older series value or a live/retained fallback. Pointer and keyboard inspection use the same single-time rule, and leaving inspection restores that default. Each panel owns its own timestamp. Only a panel with no stored observations may show explicitly qualified current, last-known or unknown-freshness fallback values. GuestDrawerHistory.readout.test.tsx verifies the default, fallback transition, zero, panel independence, inspection, refresh and access-denial boundaries; the existing pointer and sparse controls preserve dated interaction and geometry. A lone stored series renders one point at its actual timestamp and value, including measured zero, never an invented line. A group with one observation time exposes that date visibly with Single observation. No trend yet. and retains the existing accessible observation description. Multiple series with separate single observations remain separate points, inspectable at their actual times through the existing native control. Failed same-source refreshes retain these points; source/range replacement clears them along with paths and captions. GuestDrawerHistory.sparse.test.tsx checks geometry, units, provenance, invalid data, refresh and replacement. browser-tests/history-window.cjs also checks sparse/empty/current states through the production PBS drawer and direct shared renderer. This is presentation proof, not provider collection or #1723 installed acceptance. Because the plot stretches its viewBox to the panel (preserveAspectRatio="none"), single-observation dots and hover markers are zero-length, round-capped strokes with vector-effect="non-scaling-stroke", drawn in screen pixels so they stay round at any width; never a <circle> in the plot. Series lines carry data-history-series, which is how tests tell a line from a dot. A series without a reading at the inspected time reads no reading in the readout itself; there is no separate explanatory line that would appear and vanish, and move the chart, as the inspected time changes. Gridline labels state the plotted scale's value at each line: percent for utilization, °C for thermals, and a compact rate such as 5.0M/s for network and disk I/O, each centred on its SVG gridline. The I/O midpoint used to read "Avg", which was the scale midpoint, not a measured average (demand ledger "Numeric axis labels on history I/O charts", issue #1720). A panel with missing readings says so in one line: Missing observations: <series>. Lines break where a reading is missing. Shared history-chart gridlines must carry numeric labels derived from the plotted scale rather than semantic Avg / Max placeholders. Byte and byte-rate axes include their human-readable unit at each gridline, and the canvas measures both value and time labels into the same plot bounds used by geometry and hover selection so neither edge clips or drifts from the data. Related history charts that share a time range must opt into the shared hover group. The group owns one absolute hovered timestamp, while each chart maps that timestamp through its own plot geometry and nearest stored sample so crosshairs and tooltips remain time-aligned without coupling unrelated ranges or assuming identical value scales. Workload tables expose their inline history lens through one shared filter contract. Its first-use hint is visible only while bar or history metrics are available, disappears after a populated guest preview succeeds, and is passed through getWorkloadsMetricFilterProps. The provider-owned compositions ProxmoxPageSurface and VmwarePageSurface render the only WorkloadsFilter above their embedded WorkloadsSurface, which renders no toolbar of its own, and must consume that binding atomically rather than selecting display, hover, range, or hint accessors independently. Hover and range interactions remain session deduplicated so the presentation layer cannot create per-row or per-frame telemetry traffic. Provider-native inventories that do not render WorkloadsFilter are outside this guest-row contract and must not imitate only part of it under a second page-local View vocabulary. Object-detail navigation follows that same canonical split across platform and feature owners. Overview is the stable landing tab for current operational facts, while stored metric charts appear only after selecting an evidence-gated History tab rendered through frontend-modern/src/components/shared/Subtabs.tsx. The active History range belongs in the shared subtab row's trailing slot, and uses the shared filter-select presentation; storage pools, physical disks, guests, nodes, and unified resources must not render historical charts inline on Overview or reintroduce object-local tab or range-selector chrome. The metrics and chart groups may remain object-specific without changing this navigation contract.

Node history consumes the canonical metricsTarget carried by the resource projection before legacy node identifiers. Proxmox-only history uses the collector's node coordinates even when the unified resource has an agent display identity. The presence of discovery routing does not enable Agent-only disk-throughput charts. Explicit Agent linkage owns that capability distinction. nodeDrawerModel.branchcov0713.test.ts pins target precedence and unsupported target fallback, while NodeDrawer.test.tsx verifies the chart request and API-only chart groups through the rendered Overview charts.

Platform inventory presentation has one structural owner across provider pages, drawers, and inline detail rows. PlatformTableShell owns framed page tables; PlatformDetailTable, PlatformDetailTableHeader, and PlatformDetailTableBody own cardless nested tables while reusing the same header band, borders, single-line density, responsive table class, and overflow boundary. Provider-specific columns and cell contents are intentional variants, but a drawer or expanded row must not rebuild raw table / thead / tbody chrome or duplicate the shared header class strings. PlatformDetailTable also declares phoneVerticalScrollOwner="page" through the shared Table API. At the phone stage, canonical platform columns already fit their real container, so that variant clips accidental overflow instead of creating a nested scrollport that Chrome Android can stretch. Wider tables keep horizontal overflow, while overscroll-behavior-y: chain progressively removes Chromium's table-local boundary effect without blocking propagation to the app scroll shell. A platform page must not remove this variant or create a competing phone table scroller to recover columns that should be handled by responsive priority.

Expandable platform summary rows use getPlatformResourceDetailRowInteractionProps (or createPlatformResourceDetailState for their expansion state) for whole-row pointer activation and exclusion of embedded links and controls. The nested native disclosure button owns keyboard activation, focus and aria-expanded / aria-controls; static table rows are not duplicate controls. The shared TableRow installs a non-activating native click target only while an onClick action exists. WebKit touch must be able to synthesize that row's click, but actions remain document-delegated so embedded controls can stop propagation before the row action runs. Do not move the action into the native compatibility listener or add touch-end activation that opens rows during scrolling. Explicit caller-owned native listeners remain caller-owned. Table.test.tsx covers native binding, removal, single activation, child isolation and bound handlers; PulseDataGrid.test.tsx checks inheritance. The production PBS table/drawer browser fixture exercises first-tap disclosure and History in WebKit phone emulation, not installed collection or devices. A drag that selects text inside a clickable row is not a row activation. TableRow owns one createRowTextSelectionGuard() per row: it records where the row was pressed, and a click whose pointer moved more than a few pixels since that press while text inside the row is selected skips the row action. Copying a pod or container name for kubectl or docker therefore leaves the drawer, or a Proxmox node's guest selection, as it was. A press released in place still runs the action on the first click, even over text that is already selected (Chromium clears that selection only after the click), and a double-click stays two plain clicks. Custom row shells that run their own row action (GuestRow, StoragePoolRow) create the same guard and wire its mousedown handler and click check instead of writing a per-table selection test. Nested controls stop propagation and keyboard activation belongs to the disclosure button, so neither passes through the guard. Table.test.tsx and PulseDataGrid.test.tsx cover drags, in-place clicks over a selection, selections outside the row and bound handlers; PlatformResourceDetailTableRow.test.tsx covers the drawer toggle, chevron and Enter while text stays selected. The custom Workloads GuestRow shell reuses Table's exported nativeRowClickTarget marker, conditional on its own action, without inheriting shared row styling or changing alert/highlight treatment. Its name and row body must open the full guest drawer on the first touch even when the compact-layout disclosure button is visually hidden. Keep the action document-delegated; nested disclosure SVGs, external links and keyboard controls must still perform only their own action. Removing the action removes the native marker, and re-enabling it must work without replacing the row. Pointer-up and touch-end alone never perform the action. GuestRow.test.tsx covers native binding, non-activation, dynamic removal and child ordering; browser-tests/guest-row-touch.cjs covers trusted row/name taps, the production full drawer's backup/read guidance, nested controls and shared-row compatibility on Chromium desktop/phone and WebKit phone VM/CT fixtures. This is emulated input/presentation proof, not physical-device, native thaw or released acceptance. PlatformResourceDetailToggleButton is the desktop disclosure affordance and is visually removed on phone layouts where the complete row is the touch target; provider tables must not add a second mobile chevron. When row activation performs a different primary action, such as selecting a node's guests on Proxmox, the shared toggle remains visible on phones through its hideWhenRowTappableOnMobile={false} option. The node name supplies native keyboard activation for guest selection, and embedded controls retain their separate actions. Explicit guest selection focuses the guest heading with scroll prevention. The shared revealElementInViewport helper leaves a visible heading anchored and reveals an off-screen heading only far enough to show the start of its results. It uses the actual scroll container and marks deliberate movement so route-state restoration cannot compete with it. Activating the selected node again clears its node scope without moving scroll or focus, while hover leaves the guest inventory unchanged. Operator overrides remain in the shared Manage tab and use the explicit compact density of FormSelect and FormTextarea, keeping form labels, help relationships, touch targets, and control chrome canonical without expanding the low-frequency management surface. When the summary row already owns a canonical unified Resource, its expanded content composes PlatformResourceDetailTableRow instead of rebuilding a provider-local fact grid. This keeps Overview, History, Manage, accessibility, focus restoration, and responsive drawer behavior on one primitive. Proxmox Backup Server rows follow this rule and request the shared host-details disclosure open initially so a merged agent's system, hardware, network, disk, and thermal facts remain discoverable from the PBS surface. ProxmoxBackupServersTable belongs to the Proxmox Backups tab. Proxmox Overview must not duplicate that domain table between its node and guest regions; the tab boundary owns PBS server, datastore, and artifact detail. The shared Proxmox section rail renders only tabs backed by current capability evidence. It reads the source-filtered facets.byType from one compact type=pmg&source=proxmox,pbs,pmg,agent request, not estate-wide aggregations.byType: unrelated VMware VMs and TrueNAS storage must not advertise Proxmox workflows. Agent is included for Proxmox-owned physical disks, whose fact source may remain agent; because that source is shared, a generic agent disk can still expose Storage until the route filters its rows. A fresh all-resources cache does not carry source-filtered facets, so the compact query must revalidate even when that cache can paint rows. While the facets are unknown the rail must not show every optional tab as a loading fallback; an independently fetched positive replication-job count may still expose Replication. A bookmarked section remains the active hydration target until counts can distinguish unavailable from unsupported, then unsupported sections fall back to Overview. This rule applies at desktop and phone widths and must not rewrite the URL or discard a valid PBS-only Backups tab when its count arrives. When that surface receives the provider-owned PBS resource and its host Agent as separate canonical resources, ProxmoxBackupServersTable may assemble a presentation-only drawer resource only after one unique normalized host identity match. The row retains the PBS id and service facet while current host telemetry, host details, and stored History use the Agent facet and Agent metrics target. Zero or multiple matches must leave the PBS resource unchanged rather than guessing; this presentation correlation must not mutate either canonical input or create a second mobile disclosure interaction. When the PBS service has a registry-corroborated pbs.linkedAgentId, the Backups drawer selects only an Agent-bearing resource with that source-native ID, even if same-host PVE API labels or a token-auth PBS connection have no matching hostname. PVE-only rows cannot masquerade as a host series. A new link cannot reuse a retained old host target, and the displayed PBS row keeps its own canonical resource ID while History uses the selected host target.

Presentation helpers that mirror a server-side classification must name the predicate they mirror and expose it as a single exported function rather than inlining the boundary at each call site. isPhysicalDiskWearoutReported mirrors storagehealth.WearoutReported, and both the health-status and life-column helpers gate on it, so a wearout reading cannot be classified one way in a status label and another way in the cell beside it.

Feature surfaces classify resources through the shared classifier that already owns the distinction rather than re-deriving it from a raw resource type. isPulseAgentPlatformResource is the single authority for whether a resource is a standalone Pulse-agent machine or a provider-owned node, and the standalone page, platform navigation, the resource drawer and the alerts threshold sections all read it. Selecting by bare type === 'agent' collapses that distinction and puts the same machine on two surfaces that do not share an identity.

The Machines threshold tab owns one responsive SMART rules card ahead of the resource groups. It composes the existing threshold tab layout and native accessible checkbox/number controls rather than creating a second settings shell. Failed-health is presented as a toggle; sector, media, CRC-growth, remaining-life, and spare policies are integer inputs; percentage inputs are bounded to 0..100; and the card states that zero disables an individual numeric rule. Every edit uses the canonical agent-default setter and dirty-state path, so the shared save/discard bar owns persistence on desktop and narrow layouts. The single-column narrow layout must keep labels, help, values, percent suffixes, and the full scroll journey inside the app scroll shell without horizontal overflow.

The System Updates channel selector presents Stable and Preview as the two operator choices. Preview is the user-facing umbrella for governed alpha, beta, and release-candidate publications: its guidance must say that beta is for user testing while product changes are still expected, and that an RC is only appropriate when the build could become stable without product changes. The historical rc settings value remains a compatibility identifier, not a display label or permission to describe every preview as release-ready. Automatic stable updates remain unavailable while Preview is selected, and that manual-channel consequence must stay visible at desktop and narrow widths.

Backup date uncertainty uses existing age cells

Proxmox Coverage and By date age cells label unorderable restore-point dates as cautionary Unknown with fixed hover and accessible explanations, rather than using the generic relative-time formatter's future-to-now fallback. Valid ages still use the shared relative-time primitive and existing table layouts. Coverage distinguishes unknown chronology from None in each source and the independent-backup slot; expansion preserves every artifact and its separate completion/verification state. No new route or native safety claim. Production-component browser verification covers desktop and 390/320px phone layouts, scoped filtering, day selection and later inventory replacement.

Compact Coverage age cells use the existing metadata-size text and reclaim numeric-age gutters so the full Unknown word stays visible even beside a Job column at 320px; text-range checks account for cell padding, not only the outer box.

Removal copy actions are independent of install unlock state

The existing Infrastructure installer, source-manager and removed Agent Doctor controls copy removal commands before or after installation-token generation, without minting, displaying or embedding another credential. Displayed command and copied text must agree. Guidance describes saved host connection/credentials and does not present the saved Unix helper as an offline authentication bypass. The agent-uninstall.cjs parent/final desktop and touch-phone matrix exercises these production components with synthetic API data, including selected trust, canonical identity, optional authentication and installation's separate token gate. No route, navigation entry or remote-execution surface is added; native removal/failure retention and containing publication remain separate evidence.

Canonical Files

  1. frontend-modern/src/components/shared/ 1a. frontend-modern/src/components/Infrastructure/resourceDetailDrawerMetricsHistoryModel.ts 1b. frontend-modern/src/components/Workloads/nodeDrawerModel.ts 1c. frontend-modern/src/features/docker/dockerHostDrawerModel.ts 1d. frontend-modern/src/components/Workloads/NodeDrawerOverview.tsx 1e. frontend-modern/src/utils/proxmoxUpdateEvidence.ts 1f. frontend-modern/src/content/help/updates.ts
  2. frontend-modern/src/components/Settings/Settings.tsx
  3. frontend-modern/src/components/Settings/SettingsDialogs.tsx
  4. frontend-modern/src/components/Settings/SettingsPageShell.tsx
  5. frontend-modern/src/components/Settings/settingsPanelRegistry.ts
  6. frontend-modern/src/components/Settings/APIAccessPanel.tsx 6a. frontend-modern/src/components/Settings/AgentIntegrationsPanel.tsx
  7. frontend-modern/src/components/Settings/AIModelSelectionSection.tsx
  8. frontend-modern/src/components/Settings/AIProviderConfigurationSection.tsx
  9. frontend-modern/src/components/Settings/AIRuntimeControlsSection.tsx
  10. frontend-modern/src/components/Settings/AISettings.tsx
  11. frontend-modern/src/components/Settings/AISettingsDialogs.tsx
  12. frontend-modern/src/components/Settings/AISettingsStatusAndActions.tsx
  13. frontend-modern/src/components/Settings/aiSettingsModel.ts
  14. frontend-modern/src/components/Settings/AuditLogPanel.tsx
  15. frontend-modern/src/components/Settings/useAuditLogPanelState.ts
  16. frontend-modern/src/components/Settings/AuditWebhookPanel.tsx
  17. frontend-modern/src/components/Settings/useAuditWebhookPanelState.ts
  18. frontend-modern/src/components/Settings/CopyCommandBlock.tsx
  19. frontend-modern/src/components/Settings/diagnosticsModel.ts
  20. frontend-modern/src/components/Settings/DiagnosticsPanel.tsx
  21. frontend-modern/src/components/Settings/DiagnosticsResultsPanel.tsx
  22. frontend-modern/src/components/Settings/OperationsPanel.tsx
  23. frontend-modern/src/utils/diagnosticsPresentation.ts
  24. frontend-modern/src/components/Settings/GeneralSettingsPanel.tsx
  25. frontend-modern/src/components/Settings/NetworkSettingsPanel.tsx
  26. frontend-modern/src/components/Settings/RecoverySettingsPanel.tsx
  27. frontend-modern/src/components/Settings/SecurityAuthPanel.tsx
  28. frontend-modern/src/components/Settings/SecurityOverviewPanel.tsx
  29. frontend-modern/src/components/Settings/settingsHeaderMeta.ts
  30. frontend-modern/src/components/Settings/selfHostedBillingPresentation.ts
  31. frontend-modern/src/components/Settings/SSOProvidersPanel.tsx
  32. frontend-modern/src/components/Settings/useAISettingsState.ts
  33. frontend-modern/src/components/Settings/useDiagnosticsPanelState.ts
  34. frontend-modern/src/components/Settings/useSettingsShellState.ts
  35. frontend-modern/src/components/Settings/useSSOProvidersState.ts
  36. frontend-modern/src/components/Settings/ssoProvidersModel.ts
  37. frontend-modern/src/utils/ssoProviderPresentation.ts
  38. frontend-modern/src/utils/systemSettingsPresentation.ts
  39. frontend-modern/src/utils/aiSettingsPresentation.ts
  40. frontend-modern/src/utils/settingsShellPresentation.ts 42a. frontend-modern/src/i18n/
  41. frontend-modern/src/utils/textPresentation.ts
  42. frontend-modern/src/components/Settings/UpdateInstallGuide.tsx
  43. frontend-modern/src/components/Settings/updatesSettingsModel.ts
  44. frontend-modern/src/components/Settings/UpdatesSettingsPanel.tsx
  45. frontend-modern/src/components/Settings/ReportingPanel.tsx, frontend-modern/src/components/Settings/ReportScheduleEditor.tsx and frontend-modern/src/components/Settings/ReportingFormField.tsx
  46. frontend-modern/src/components/Settings/reportingPanelModel.ts 48a. frontend-modern/src/components/Settings/reportingSchedulesModel.ts
  47. frontend-modern/src/components/Settings/reportingInventoryExportModel.ts
  48. frontend-modern/src/components/Settings/useReportingPanelState.ts
  49. frontend-modern/src/utils/reportingPresentation.ts
  50. frontend-modern/src/utils/updatesPresentation.ts
  51. frontend-modern/src/components/Settings/__tests__/settingsArchitecture.test.ts
  52. tests/integration/tests/15-settings-shell-consistency.spec.ts
  53. frontend-modern/src/components/shared/FilterBar/FilterBar.tsx
  54. frontend-modern/src/components/shared/FilterBar/FilterChip.tsx
  55. frontend-modern/src/components/shared/FilterBar/AddFilterMenu.tsx
  56. frontend-modern/src/components/shared/FilterBar/filterCatalog.ts
  57. frontend-modern/src/components/shared/FilterBar/index.ts 59c. frontend-modern/src/components/shared/FilterBar/filterOptionPresentation.tsx
  58. frontend-modern/src/components/shared/TypeColumn.guardrails.test.ts
  59. frontend-modern/src/features/ (including Patrol presentation, where Patrol control starter counts are context only even when mirrored through patrolAutonomy* compatibility fields, successful direct Patrol control saves may record the content-free patrol_control starter only when paid control is available and the effective control posture changes and must then refresh Patrol status, findings, approvals, and run history before the operator waits for polling, legacy proActivation* starter aliases must not render a separate proof strip by themselves, Patrol control completed/resolved counts may only project backend-owned terminal proof, current active findings and pending approvals outrank historical completion proof in the primary operator state, selected run history must read as a Patrol run record rather than a findings filter or snapshot workflow, terminal verified/rejected outcomes with no active finding or pending approval must stay history detail without rendering a no-op proof strip, resolved-only issue history must not be promoted into current-work copy or actions, compact recurrence/trust counters must read as historical evidence rather than current issue state, Patrol-owned status/history evidence must keep the assessment visible when the broader intelligence summary is missing, Patrol work-group chips may group current approvals, failed actions, failed checks, recurring active issues, and stale scheduled protection but must not become a separate status/trust/proof strip, the Patrol route and page title must lead with Patrol while the default workspace underneath may use Open work and run history stays a deliberate secondary review surface, setup-only Patrol runtime failures must instead use Fix Patrol setup framing with a dedicated setup task and direct provider-settings action while suppressing generic issue-row chips and filter chrome, Patrol must not expose a generic Details/supporting-context panel for nearby activity, related patterns, or policy limits, locked-control copy must state the watch-only boundary in positive capability language by saying Patrol checks infrastructure and shows current issues, avoid repeating the same sentence across the header and control, and avoid repeatedly restating infrastructure-unchanged caveats or relying on disabled controls, compact Pro badges, Limits controls, or manual-review framing, patrolControlValueState decides whether a terminal decision is partial review context or verified value proof while patrolAutonomyValueState remains a compatibility mirror, legacy proActivation* fields are compatibility fallback only, native Patrol state must not load the operations-loop status projection to decide current work, local Patrol state must expose issue-backed patrolWork* evidence rather than legacy proof naming, Patrol mode labels remain domain copy that must describe backend-owned risk policy without creating page-local safety thresholds, the selected Patrol mode sentence must state the approval and policy boundary without adding a second limits panel or proof strip, Patrol schedule, trigger, and model settings must stay on Settings > Pulse Intelligence > Patrol, apart from the Patrol mode selector, instead of returning as a Patrol page drawer, that settings page must not duplicate the four control choices, and the Patrol page mode selector must not reintroduce save/apply configuration framing, and Patrol header refresh controls must call an explicit operator-refresh handler whose spinning/disabled state is separate from background polling and initial data loads)
  60. frontend-modern/src/components/SetupWizard/SetupWizard.tsx
  61. frontend-modern/src/components/Settings/useSettingsInfrastructurePanelProps.ts
  62. frontend-modern/src/components/SetupWizard/SetupCompletionPreview.tsx
  63. frontend-modern/src/components/SetupWizard/steps/WelcomeStep.tsx
  64. frontend-modern/src/components/SetupWizard/__tests__/SetupWizard.test.tsx
  65. frontend-modern/src/components/SetupWizard/__tests__/SetupCompletionPreview.test.tsx
  66. frontend-modern/src/components/SetupWizard/__tests__/WelcomeStep.test.tsx
  67. frontend-modern/src/components/Settings/SystemLogsPanel.tsx
  68. frontend-modern/src/components/Settings/useSystemLogsPanelState.ts
  69. frontend-modern/src/utils/systemLogsPresentation.ts
  70. frontend-modern/src/components/Settings/__tests__/SystemLogsPanel.test.tsx
  71. frontend-modern/src/components/Settings/ResourcePicker.tsx
  72. frontend-modern/src/utils/reportableResourceTypes.ts
  73. frontend-modern/src/utils/reportingResourceTypes.ts
  74. frontend-modern/src/utils/workloadEmptyStatePresentation.ts
  75. frontend-modern/src/utils/workloadGuestPresentation.ts
  76. frontend-modern/src/utils/emptyStatePresentation.ts
  77. frontend-modern/src/utils/semanticTonePresentation.ts
  78. frontend-modern/src/components/Toast/Toast.tsx
  79. frontend-modern/src/utils/toast.ts
  80. frontend-modern/src/utils/semanticTonePresentation.ts
  81. frontend-modern/src/utils/emptyStatePresentation.ts
  82. frontend-modern/src/utils/typeColumnPresentation.ts
  83. frontend-modern/src/components/Settings/NetworkBoundarySettingsSection.tsx
  84. frontend-modern/src/components/Settings/networkSettingsModel.ts
  85. frontend-modern/src/components/Settings/useDiscoverySettingsState.ts
  86. frontend-modern/src/components/Settings/useSettingsInfrastructurePanelProps.ts
  87. frontend-modern/src/components/Settings/AvailabilitySettingsPanel.tsx
  88. frontend-modern/src/components/Settings/availabilitySettingsModel.ts
  89. frontend-modern/src/components/Settings/settingsPanelRegistryContext.tsx
  90. frontend-modern/src/components/Settings/settingsPanelRegistryLoaders.ts
  91. frontend-modern/src/components/Settings/settingsNavigationModel.ts
  92. frontend-modern/src/components/Settings/settingsNavCatalog.ts
  93. frontend-modern/src/components/Settings/settingsNavVisibility.ts
  94. frontend-modern/src/components/Settings/settingsRouting.ts
  95. frontend-modern/src/components/Settings/settingsTabSaveBehavior.ts
  96. frontend-modern/src/components/Settings/settingsTypes.ts
  97. frontend-modern/src/components/Settings/useSettingsNavigation.ts
  98. frontend-modern/src/components/Settings/useSettingsPanelRegistry.tsx
  99. frontend-modern/src/components/Settings/useSettingsSystemPanels.tsx
  100. frontend-modern/src/components/Settings/DockerRuntimeSettingsCard.tsx
  101. frontend-modern/src/components/shared/EnvironmentLockBadge.tsx
  102. frontend-modern/src/utils/environmentLockPresentation.ts
  103. frontend-modern/src/utils/docsLinks.ts
  104. tests/integration/tests/20-local-doc-links.spec.ts
  105. frontend-modern/src/index.css
  106. frontend-modern/src/components/shared/summaryInteractionA11y.ts
  107. frontend-modern/src/components/shared/SummaryRowActionButton.tsx
  108. frontend-modern/src/hooks/createNonSuspendingQuery.ts Shared API read-access classification is owned by frontend-modern/src/utils/apiAccessError.ts; it recognises only final numeric 401/403 statuses and never renders transport diagnostics. 111a. frontend-modern/src/utils/storageSummaryCache.ts
  109. frontend-modern/src/components/shared/TableCardHeader.tsx
  110. frontend-modern/src/components/shared/UpgradeLink.tsx
  111. frontend-modern/src/utils/upgradeNavigation.ts
  112. frontend-modern/src/components/DemoBanner.tsx 116a. frontend-modern/src/components/CommercialMigrationBanner.tsx 116b. frontend-modern/src/components/GitHubStarBanner.tsx
  113. frontend-modern/src/components/Login.tsx
  114. frontend-modern/src/stores/sessionCapabilities.ts
  115. frontend-modern/src/stores/sessionPresentationPolicy.ts
  116. frontend-modern/src/stores/licenseCommercial.ts
  117. frontend-modern/src/useAppRuntimeState.ts
  118. frontend-modern/src/routing/routePreload.ts
  119. frontend-modern/src/stores/aiAvailability.ts
  120. frontend-modern/scripts/header-audit.mjs
  121. frontend-modern/src/components/Settings/DataHandlingPanel.tsx
  122. frontend-modern/src/components/Settings/dataHandlingPanelModel.ts
  123. frontend-modern/scripts/canonical-platform-audit.mjs
  124. frontend-modern/scripts/settings-diagnostics-boundary-audit.mjs
  125. frontend-modern/scripts/shared-template-audit.mjs
  126. frontend-modern/scripts/shared-template-registry.json 129a. frontend-modern/scripts/planning-doc-status-audit.mjs 129b. frontend-modern/scripts/__tests__/planning-doc-status-audit.test.mjs
  127. frontend-modern/src/features/platformPage/sharedPlatformPage.tsx 131a. frontend-modern/src/features/platformPage/platformSearchSuggestions.ts 131b. frontend-modern/src/features/platformPage/PlatformResourceDetailTableRow.tsx 131c. frontend-modern/src/features/platformPage/PlatformOutdatedAgentNotice.tsx 131d. frontend-modern/src/features/platformPage/PlatformOutdatedSensorSetupNotice.tsx 131e. frontend-modern/src/features/platformPage/platformEstateOverviewModel.ts 131f. frontend-modern/src/components/shared/Form.ts 131g. frontend-modern/src/components/shared/FormSelect.tsx 131h. frontend-modern/src/components/shared/FormTextarea.tsx 131i. frontend-modern/src/components/Infrastructure/ResourceOperatorStateSection.tsx
  128. frontend-modern/src/utils/platformSupportManifest.generated.ts
  129. frontend-modern/src/utils/platformSupportManifest.ts
  130. frontend-modern/src/utils/sourcePlatformOptions.ts
  131. frontend-modern/src/utils/sourcePlatforms.ts
  132. frontend-modern/src/utils/infrastructureOnboardingPresentation.ts
  133. frontend-modern/src/components/shared/Button.tsx
  134. frontend-modern/src/components/shared/buttonModel.ts
  135. frontend-modern/src/components/shared/Button.test.tsx 139a. frontend-modern/src/components/shared/InlineNotice.tsx 139b. frontend-modern/src/components/shared/InlineNotice.test.tsx 139c. frontend-modern/src/components/shared/ExternalTextLink.tsx 139d. frontend-modern/src/components/shared/ExternalTextLink.test.tsx
  136. frontend-modern/src/components/shared/DetailSectionTable.tsx
  137. frontend-modern/src/components/shared/detailSectionModel.ts
  138. frontend-modern/src/components/Settings/__tests__/settingsLocalization.test.ts
  139. frontend-modern/src/i18n/__tests__/i18n.test.ts
  • frontend-modern/src/components/Settings/BackupTransferDialogs.tsx
  • frontend-modern/src/components/Settings/useBackupTransferFlow.ts

Shared Boundaries

Transport-independent operator rows

Proxmox/Storage and Docker presenters retain the shared table/detail, responsive column weighting, relative-time and governed action primitives. Display names for Proxmox storage plugins preserve their operator vocabulary (LVM-Thin, NFS, SMB/CIFS and the other supported acronyms); only explicitly shared storage on a synthetic cluster/shared location gets the shared label. That label does not alter canonical identity, parent, History target or pool-to-disk ownership.

The Docker image phone projection pairs hidden headers and cells and gives measured size and update verdict room without hiding identity or dropping the existing detail disclosure. The container projection retains its shared 30% phone identity reservation and five scan signals; off-row restart/host/consumer evidence remains in detail. Swarm rollout tokens read as words, with compact Running headings and rollout states that truncate on one line on phones; raw rollout and placement evidence remain reachable through the title and keyboard/touch row disclosure. platformOverviewLayout.guardrails.test.ts, mounted Docker table tests and the desktop/touch-phone production-component browser fixture check these priorities, rather than treating an old percentage as an immutable primitive contract. No new public route or navigation entry is added.

Maintenance schedule copy

Operator maintenance banners show local absolute start and end timestamps, including future dates. Past-time relative formatting must not turn a future expiry into "just now". The banner describes paused attention rather than promising that rejected alert observations are acknowledged. Active and future-window mounted regressions pin these user-visible claims.

Platform tables read staleness from one helper

features/platformPage/resourceStaleness.ts turns the backend's stale health verdict into a short label for platform tables, so a table that stops presenting a silent resource's last metrics as live does it the same way: dimmed metric cells plus the age with a no-signal icon, the full sentence on hover and as the accessible name. Grouped platform tables that already name the group in a shared grouped-table header row drop columns that would repeat the group's own value on every row.

Ceph health projections

ProxmoxCephTable uses one health classification for the row, status-filter counts and drawer header. Native HEALTH_ERR and the server-normalized ERR both remain critical, including when no manager daemons are reported. A missing service tally must not turn explicit critical provider health into a healthy table verdict. Mounted regressions in ProxmoxCephTable.test.tsx pin the native/normalized vocabulary and selected/unselected critical filtering.

Retained state in bounded platform windows

PlatformWindowedRows and PlatformWindowedList keep one keyed renderer owner while the visible window moves. Items present in both windows retain component identity, active detail tabs and unsaved edits. Passing a new window must update the existing renderer rather than instantiate another renderer around it. Removed items still unmount, and row budgets and spacer behavior are unchanged. The mounted scroll regression in PlatformWindowedRows.test.tsx exercises both renderers with an edited input retained across overlapping windows. The browser journey is a narrow Proxmox node Manage form, scrolling to lifecycle Save while the synthetic estate continues updating, plus the alert timeline and note form.

PBS host history correlation

The Backups surface passes its complete deduplicated route inventory to the PBS table. Only PBS servers render as rows; other resources supply correlation inputs. Preserve PBS drawer identity and use the correlated canonical history target. Missing disk utilisation does not gate CPU/memory history.

One agent can be surfaced twice for a single PBS host: folded into its PVE guest and as a standalone source=pbs host row. Those two rows are one machine, not an ambiguous pair. Correlation must collapse candidates that share an agent identity and prefer the guest representation, whose canonical metrics target carries the persisted host history; the PBS service target has no stored host series and renders the no-history state. Two candidates with distinct agent identities remain ambiguous, and a candidate with no agent identity must not be treated as proof of sameness.

A live snapshot can briefly omit the correlated host row while the PBS server row remains, for example while a realtime refresh replaces the merged estate. The correlation must retain the last resolved host per PBS server across that omission instead of falling back to the PBS service target, so the drawer's linked host and History series do not flicker. Reuse the remembered host only while it is still fresh relative to the server, and drop it once stale so a removed or replaced host is not advertised indefinitely; a host row that is present but ambiguous still declines.

Verification: ProxmoxBackupServersTable.drawer.test.tsx covers standalone and merged guest targets, missing disks, duplicate guest/host representations of one agent, genuinely ambiguous identities, and retaining the resolved host target across a transient host-row omission; ProxmoxBackupServersTable.test.ts covers retention, staleness and pruning; ProxmoxPageSurface.contract.test.tsx covers hydration and deduplication.

The settings panel registry supplies organisation overview, access and sharing with security-status currentUsername. An explicitly empty principal must not fall back to configured administrator identity; only older responses lacking the field use proxy/SSO/configured-username compatibility fallback. This identity plumbing must not change settings shell framing or bypass panel capability gates.

Settings navigation discoverability is part of the shared settings-shell boundary. A settings route that is available in normal commercial presentation must be reachable through the sidebar unless it is explicitly a hidden deep-link flow. system-billing (Plans & Billing) is listed for every session that does not hide commercial surfaces, including ordinary free self-hosted sessions, so an operator can find the plan and buy; only demo and white-label runtimes hide it. Ordinary free sessions keep the quiet boundary for everything proactive: paid-feature items including system-relay, support-reporting, security-roles, security-users, security-audit, and security-webhooks use hideWhenUnavailable. Their direct routes remain available because the owning panels still handle explicit activation, recovery, and feature-gate handoffs. Capability-based hiding still applies independently, so a session that lacks route authority cannot mount a panel merely because it has paid or recovery context.

Candidate import-plan presentation inside the Infrastructure settings dialog is a shared primitive composition boundary. NodeCandidateImportPlan.tsx may use shared Button, checkbox styling, lucide icons, and MonitoredSystemImpactPreview, while InfrastructureWorkspace.tsx owns the route-backed dialog state that feeds probe or Discovery candidates into the credential slot. That surface must keep the approval card readable inside the existing dialog body and must not introduce a second nested modal, detached wizard shell, or page-local preview renderer for monitored-system impact.

Frontend localization is a shared primitive boundary. Locale support must flow through typed message catalogs with an English fallback and explicit seed locale coverage rather than page-local string switches. Alert snooze and resume copy follows that same boundary: action labels, bounded duration presets, monitoring-continuity wording, exact-expiry presentation, failure feedback, and timeline labels must stay in the typed English, German, Spanish, and French catalogs. The product surface may format the chosen instant in the viewer locale, but must not duplicate those lifecycle strings inside the alert card or dialog component.

French is a first-wave locale alongside German and Spanish: a lazily loaded frontend-modern/src/i18n/messages.fr.ts must cover every migrated key, fr-* browser languages resolve to it, and the Settings > General language selector lists every entry of SUPPORTED_LOCALES from the registry rather than a hand-written option list. Because that selector grows with the registry, its row stacks the label above the options below the sm breakpoint so every locale stays on one line at phone width instead of clipping the last option. Count-bearing catalog keys have no plural forms, so locales whose nouns or adjectives agree with the number must use count-neutral wording (for example Critique : {count}) rather than a form that is wrong for one or for many.

Date and number formatting is part of that boundary and is separate from the message catalog. Every toLocaleString, toLocaleDateString, toLocaleTimeString, and Intl.* constructor must pass undefined as the locale so the runtime resolves the viewer's own locale, including whether they expect a 12 or 24 hour clock and which order the date parts go in. A literal tag such as 'en-US' shows US conventions to every reader everywhere. The canonical-shared/no-hardcoded-format-locale rule in frontend-modern/scripts/canonical-platform-audit.mjs blocks new occurrences; a call that genuinely needs a fixed locale, such as a stable machine-readable export, belongs in that rule's allowFiles with a reason rather than as a bare literal. The rule exists because this regressed silently once already: #1279 delocalized the "Last refresh" clock in App.tsx, and the v6 rewrite that moved that logic into frontend-modern/src/useAppRuntimeState.ts reinstated the hardcoded form, which shipped through v6 GA until #1685. frontend-modern/src/i18n/locales.ts owns locale normalization, the supported locale registry, and fallback chains; frontend-modern/src/i18n/messages.ts owns the typed catalog shape; and frontend-modern/src/i18n/policy.ts owns the first-wave non-translatable token rules. The active app locale is a shared user preference initialized from stored or browser language and exposed through Settings > General; individual surfaces must consume that shared preference instead of creating local language toggles. Customer-facing shell, navigation, settings, first-run, empty-state, commercial handoff, and alert copy may be localized through this catalog — including the alert-to-Patrol action surface ("Have Patrol investigate" and its targeted-check menu hint) that is primary on resource-backed active alert cards, plus the secondary Assistant explanation handoff strings — but machine-facing values must remain stable: commands, environment variables, API fields, config keys, log lines, error codes, hostnames, resource names, product identifiers, and vendor object names stay untranslated unless the owning runtime contract explicitly says otherwise. Shared settings-shell header copy for the self-hosted plan must keep first-wave locale catalogs aligned with the English product stance: on Pro, the operator chooses how autonomous Patrol should be. The settings shell must not teach a separate activation loop, MCP readiness, or operations-loop proof model as the default plan setup story. Pulse Intelligence external-agent setup uses the same shell language with a Choose Patrol mode handoff before scoped-token setup, and the expanded setup checklist must say to set how autonomous Patrol should be before connected agents request work rather than repeating internal automation/proof wording. Developer-only external-agent posture uses External agents plus Patrol mode before surfacing MCP or workflow prompt wire names. Migrated settings surfaces must render customer-facing copy through the catalog and shared presentation helpers rather than reintroducing panel-local English. Migrated first-session surfaces, including the Setup Wizard shell, welcome/security steps, setup completion handoff, and runtime-home loading handoff, follow the same catalog path; their guardrails must fail if the migrated journey returns to page-local English or translates commands, URLs, generated credentials, product/source identifiers, or reported resource names. Migrated Alerts Overview surfaces, including the page shell, overview stat cards, active-alert triage list, acknowledgement actions, incident timeline panel/filter controls, and Pulse Assistant alert handoff briefing, must also route user-visible copy through the catalog and alert-owned presentation helpers. Alert IDs, alert types, resource IDs, resource names, node names, source messages, event payloads, commands, command output, logs, and Assistant model-context labels stay machine-stable and untranslated. The Alerts Overview hydration boundary must localize its pending, unavailable, and retry presentation through the same catalog. Pending and unavailable copy must explicitly withhold an all-clear until active-alert truth is confirmed; the surface must not reuse the localized zero-alert empty state while the canonical alert snapshot is unknown. First-wave catalogs and their focused catalog proof must advance together so this reliability distinction cannot silently fall back to English or collapse into ordinary empty-state wording. The active-alert card's delivery-status row remains an alerts-owned compact presentation composed inside the shared responsive card shell. It consumes one bulk diagnosis snapshot per overview refresh, wraps beside started-at and threshold metadata at narrow widths, and does not create a second card or page-level banner. Acknowledged cards omit the redundant row, and unavailable diagnosis data leaves the existing alert card intact. The Destinations delivery-activity surface follows the same ownership split: alerts own the AlertDeliveryLogCard composition inside the shared Card, button, badge, list, and responsive wrapping primitives. Its single newest-first list may mix notification delivery attempts with alert-policy held events, but each row retains its owning outcome vocabulary and evidence tooltip. At narrow widths, resource, reason, and relative-time fields wrap or truncate inside the card without creating page-level horizontal overflow, and the shared refresh action remains independently usable while the bounded list scrolls. The legacy pricing handoff page may also route its visible redirect title and manual-link copy through the catalog, but Pulse Account, route paths, feature keys, query parameters, public URLs, and purchase-return state remain machine-stable and untranslated. Customer-facing copy must use complete sentences or an intentional visual separator such as a middle dot. Semicolons are reserved for machine-facing syntax including commands, cookies, encoded data, CSS declarations, and HTML entities. frontend-modern/scripts/copy-style-audit.mjs enforces that boundary across production TypeScript and TSX, and every ordinary product-copy exception must be rewritten instead of allowlisted.

Alert thresholds consume the shared FilterBar primitive and route state, while the alerts subsystem owns the resource data and platform-specific threshold tabs. The thresholds platform IA is platform-shaped: Proxmox, Docker, Kubernetes, TrueNAS, vSphere, PBS, PMG, and Systems. Frontend primitives own the chip, reset, "+ Filter", and route-backed shell pattern; alerts must not replace that with page-local search/tab chrome or legacy neutral buckets. Threshold edit semantics under frontend-modern/src/features/alerts/thresholds/ also stay alerts-owned: the override mutation hooks write sparse enabled-only backup/snapshot overrides that inherit global thresholds at evaluation time, and the warning/critical pair reconciliation in the thresholds helpers adjusts the untouched field on a conflicting single-field edit. Primitives must not absorb those persistence or validation rules into shared form/table chrome (#1126).

The alerts-owned threshold surface also owns rolling metric evaluation configuration. ThresholdsTab presents the understandable policy (current value or a named rolling duration) rather than a query language, with one global CPU rule, a canonical all-workload fallback, and sparse platform overrides that visibly inherit the effective parent. The select options and inherit labels must mirror the runtime hierarchy: VMs and application containers inherit guest, agent-backed systems inherit node, and every chain terminates at all. The configuration snapshot and payload models must round-trip explicit zero and resource-specific maps without flattening inheritance. Copy must explain that incomplete history holds incident state instead of firing or recovering, and the control must remain usable at desktop and narrow viewports through shared Card, label, and native-select primitives. PlatformTableToolbar may accept compact consumer-owned context through its shared leading-control rail when the context is actionable for the inventory immediately below it. Consumers must use that extension point for local attention counts and review actions instead of introducing a parallel filter card or a routine page-wide posture banner; the toolbar continues to own responsive placement alongside View preferences and row counts. Frequently used leading context remains visible when the mobile filter body is collapsed and moves into the canonical action rail when that body is expanded; lower-priority trailing orientation controls remain collapsed with the rail. Native multiline form fields are also a shared primitive boundary. FormTextarea owns label/id/help wiring, controlled value synchronization, and textarea chrome for alert, settings, and infrastructure runtime surfaces; those surfaces must not recreate raw native <textarea> shells locally.

Toast notification chrome is a shared primitive boundary. Toast owns the global notification stack shell, status icon placement, and dismiss action chrome; status glyphs must come from the shared library icon set and dismiss controls must compose ActionIconButton. Consumer-specific raw SVG status icons, toast-local close-button class strings, and page-local toast stacks are forbidden unless the shared-template registry is intentionally extended first.

The System Updates install guide is a shared settings primitive, not a deployment-lane-only panel. UpdateInstallGuide must render the canonical update-plan readiness verdict inline with the update action, and a blocked readiness status must make automatic install visibly unavailable until the blocking check is resolved. That install guide also owns the Pro-runtime Docker guidance: when the compiled runtime identity is pro (runtimeCapabilities().runtime.build, the same signal the update banner keys off), the Docker install steps and the idle Docker box must render the license server broker's digest-pinned commands from UpdateInfo.dockerUpdate (or an explicit Pro notice when they are absent) and must never show the community rcourtman/pulse pull commands, which would silently downgrade the container to the community build.

Platform page subnavigation is a shared frontend primitive. Docker / Podman and Kubernetes platform pages may add native API-backed sections, but the tabs must use PlatformSectionTabs, canonical table alignment helpers, and shared resource type presentation/reporting helpers rather than page-local tab shells, alignment classes, or ad hoc report-category coercion. Platform tabs are responsively scrollable within that shared shell. When the rail overflows, the primitive must expose focusable, labeled previous/next controls and keep their availability synchronized with the rail position so hidden workflows remain discoverable without widening the document or requiring precision swipes. At the leading edge, the first tab must begin flush with the rail instead of reserving an empty previous-control slot; trailing space may be reserved only while the next-control affordance is actually present. Platform tabs are feature-owned consumers of canonical resource payloads: Docker page model helpers may prefer backend-authored DockerData stack and Podman metadata for search/display while shared primitives continue to own only the tab shell, filter controls, and reusable presentation affordances. Platform tabs are workflow-level navigation, not one visible tab per API resource kind, and they must be evidence-gated by their owning row or signal model. Overview is the stable landing surface; supporting workflow tabs appear only when the current setup has native inventory or signal for that workflow, and legacy object URLs resolve to their owning workflow only when that workflow is visible. Docker / Podman may expose Overview, Images, Storage, Networks, and Swarm, while legacy /docker/containers resolves to the Overview landing surface rather than remaining a separate visible tab. Kubernetes may expose Overview, Nodes, Workloads, Services, Storage, Configuration, and Events; TrueNAS and vSphere follow the same evidence-gated primitive for native storage, service, app, VM, protection, datastore, network, health, and activity workflows. API-native tables remain bespoke under those workflows, so Docker Overview owns runtime hosts plus primary container workloads, Docker Storage owns engine disk usage plus volumes, and Docker Swarm owns services, tasks, nodes, secrets, and configs; Kubernetes Workloads owns Pods, Deployments, controllers, and autoscaling, Services owns Services plus ingress/endpoint inventory, and Configuration owns config plus policy inventory. Backup and recovery platform pages follow the same navigation primitive boundary: source-specific evidence tables may be exposed as secondary drilldowns under an owning workflow tab, but the shared tab shell must not grow one top-level tab for each API source merely because that source has a table. Legacy object-specific URLs may resolve to the owning workflow tab, but they must not reappear as top-level platform navigation unless the product IA is intentionally changed. Overview tabs must stay deliberately shaped around the primary operator job instead of repeating every detail table:

The Proxmox backup workflow follows this same boundary: its chronological date feed and workload coverage posture are route-backed sections under /proxmox/backups, and their navigation must compose PlatformSectionTabs instead of a page-local segmented control or query-only view switch.

Docker / Podman Overview owns runtime hosts and primary container workloads in the proven host-then-workloads pattern, while Kubernetes Overview owns cluster/control-plane rollup; supporting object tables live in their dedicated workflow tabs. Docker / Podman native subsections now include runtime containers, engine storage usage, Swarm node inventory, and metadata-only Swarm secret/config inventory where the documented Docker APIs report those resources; Podman-only libpod pod inventory must not be represented until the collector has a libpod-native source. Kubernetes config inventory must preserve the same trust boundary for metadata-only ConfigMap and Secret rows: the shared table wording may indicate metadata-only inventory, but must not imply payload fields were read when the agent used Kubernetes metadata-only API responses, and the unified-resource owner supplies the Namespace, ConfigMap, Secret, and ServiceAccount-specific columns. Kubernetes Node inventory must also be reachable through a dedicated native tab, not only the overview stack, while retaining the shared PlatformSectionTabs shell. Primary app-shell navigation consumes unified-resources-owned resource evidence: empty or generic compatibility facets do not admit runtime-lens tabs on their own, and cached resource evidence must not render primary platform navigation before the first authoritative resource snapshot resolves. A platform that is not admitted stays reachable by direct setup URL, but does not occupy primary navigation with an empty page. Feature-owned Docker / Podman action controls may render backend actionReadiness disabled reasons, but the shared primitive layer owns only the button/table affordance shell; it must not invent action availability, command agent state, or alternate execution routes. Kubernetes Service inventory must likewise stay on the shared tab, toolbar, table, table-alignment, and inline-detail primitives while the unified-resource owner supplies Service type, virtual IP, published port, and selector columns. Kubernetes storage inventory follows that same primitive boundary while the unified-resource owner supplies StorageClass, PersistentVolume, and PersistentVolumeClaim-specific columns. Kubernetes networking inventory also follows that same primitive boundary while the unified-resource owner supplies Service, Ingress, and EndpointSlice-specific columns. Outdated-agent notices on platform pages are part of this same shared frontend/platform primitive boundary and must compose frontend-modern/src/components/shared/InlineNotice.tsx for the dense notice shell, icon/content layout, and action-link chrome. Agent-backed and hybrid platform pages may surface a compact stale-agent cue when their row model carries Pulse agent identity and version evidence, but the CTA must route to the canonical Infrastructure settings update-command surface with scoped agent IDs instead of duplicating installer command assembly, tokens, or lifecycle copy in each platform page. These notices must compare against the API-owned agentUpdateTargetVersion rather than the app build version, so development builds can show their dirty server version without implying agents can or should update toward that build. Agentless API-only platforms such as vSphere must not grow this notice unless a concrete guest or monitored system row actually carries a Pulse agent identity. On vSphere specifically, stale-agent copy belongs to correlated in-guest VM agents and must not describe ESXi hosts as Pulse-agent update targets just because phase-1 host resources use the canonical agent resource type. Kubernetes is cluster-agent-backed: canonical k8s-node rows may be pure Kubernetes API rows rather than merged agent rows, so the unified-resource owner must project the cluster agent identity and cluster-scoped agent version onto those node rows before the shared stale-agent collector can decide whether the node inventory is gated by an older agent. The Docker duplicate-identity warning follows the same shared notice boundary: it composes InlineNotice, is driven only by server-authored DockerData.identityConflict evidence (never by page-local hostname comparison), and names the flapping hostnames with the machine-id remedy in the notice copy because the fix happens on the operator's machines, not on a Pulse settings surface, so it carries no action link. Unlike stale-agent cues it is a data-reliability warning and is not gated on the read-only presentation policy. Global dismissible notice bars are also part of the shared InlineNotice boundary. frontend-modern/src/components/DemoBanner.tsx and frontend-modern/src/components/CommercialMigrationBanner.tsx must compose InlineNotice with the banner layout, the shared icon library, action slots, and the primitive's dismiss slot instead of carrying local colored notice shells, raw SVG status icons, or page-local action and close button classes. Kubernetes policy inventory follows that same primitive boundary while the unified-resource owner supplies NetworkPolicy policy type and rule-count columns, PodDisruptionBudget budget and observed health columns, ResourceQuota hard/used quota columns, and LimitRange item-type columns. Kubernetes autoscaling inventory follows that same primitive boundary while the unified-resource owner supplies HorizontalPodAutoscaler scale target, replica bounds, current/desired replicas, and metric source columns. Kubernetes events inventory follows that same primitive boundary while the unified-resource owner supplies Event type, reason, involved-object, count, observed-time, and message columns. Docker empty-state guidance on platform pages follows the same shared platform primitive boundary: it may use the route-specific Docker / Podman vocabulary, but it must distinguish standalone Docker host installation from the Proxmox LXC Docker host-side inventory path without adding page-local installer command assembly or token handling. The empty state must provide a direct action into the Docker-specific Infrastructure add flow rather than leaving the operator at a descriptive dead end. Docker / Podman inventory follows that same primitive boundary while the unified-resource owner supplies API-object-specific container, image, volume, network, Swarm node, task, secret, and config columns through dedicated native tables. The Docker containers tab must use the native DockerContainersTable for container state, health, restart, image, port, network, mount, update, governed lifecycle actions, and host/runtime columns rather than embedding WorkloadsSurface. The lifecycle action column is a compact icon-button primitive over unified-resource capabilities and the shared resource-action API client; it must not grow Docker/Podman shell, SSH, or provider calls inside the table component. The same governed lifecycle controls may appear in the resource detail header, while the platform table and drawer shells remain presentation owners only: they may pass a post-success refresh callback to their existing resource query, but must not own execution, approval, policy, or provider dispatch. Swarm services must surface the API-reported rollout/update state in the native services table, and engine storage rows must expose a stable row hook so platform-page browser proof can verify the storage tab is hydrated from runtime disk-usage data. Kubernetes deployments must surface the API-native observed generation and metadata age columns through the shared table shell without page-local alignment helpers or generic infrastructure columns. Docker network rows use the same shared table/detail primitive split: the default table columns prioritize attached workloads, attention state, subnets, driver, and host, while lower-level scope, addressing, flags, and network id details belong in the inline row disclosure. Attached container names, network addresses, image, health/state, and published ports are feature-owned data, but search and disclosure behavior must remain inside the shared table chrome rather than a card deck, nested card, or route-changing object browser. Dense networks must keep the inline disclosure bounded by default and provide local attached-container search, status grouping, and attention/running/other filters so large bridge or overlay networks remain scan-friendly without hiding any container from drilldown.

  1. frontend-modern/src/components/CommercialMigrationBanner.tsx shared with cloud-paid: the global commercial migration notice is both a cloud-paid entitlement recovery surface and a shared app-shell notice primitive consumer.
  2. frontend-modern/src/components/Infrastructure/useTableWindowing.ts shared with performance-and-scalability: the shared bounded table-window controller is both a canonical frontend rendering primitive and a fleet-scale scrolling hot-path boundary.
  3. frontend-modern/src/components/Settings/AgentIntegrationsPanel.tsx shared with ai-runtime, api-contracts: the External agents settings panel is the optional settings-shell projection of Pulse MCP onboarding, the AI runtime connected-agent onboarding surface, and a presentation consumer of the shared agent capabilities frontend client.
  4. frontend-modern/src/components/Settings/APIAccessPanel.tsx shared with security-privacy: the API Access settings intro is both a security/privacy token-management trust surface and a canonical settings-shell presentation boundary. The panel may own shell placement and local action layout, but token-specific Docker / Podman copy must come from frontend-modern/src/utils/apiTokenPresentation.ts rather than page-local text. frontend-modern/src/components/Settings/AgentIntegrationsPanel.tsx stays under Pulse Intelligence Assistant settings because connected agents are an optional access path to Patrol work, while API Access remains the token minting surface linked from setup. Its setup copy must present one shared pulse-mcp runtime contract with client-native wrappers: OpenCode's top-level opencode.json / mcp shape and the common mcpServers shape for Claude-style clients. The server name, command, base URL flag/default, token environment variable, supported config families, and copied config snippets must flow through frontend-modern/src/api/agentCapabilities.ts from /api/agent/capabilities.mcpAdapter; the component may arrange and label them, but must keep setup mechanics, raw client config snippets, and developer details behind deliberate disclosures so the default Assistant settings view stays focused on chat command access and optional external access rather than raw JSON. External-agent posture should present External agents as the visible surface and reserve Pulse MCP or pulse_operations_loop for wire-name/debugging detail. The Patrol control handoff must route to the Patrol operator surface where the inline control level is configured. When setup is opened, the setup order must put Patrol control before scoped-token creation and client connection, while installer commands, client config snippets, and developer details remain deliberate disclosures. It must not frame Pulse MCP as a Claude-only surface, force OpenCode through a Claude-style wrapper, or duplicate a client-specific tool list. Full-surface token guidance in that panel must render the manifest-provided requiredScopes list through frontend-modern/src/api/agentCapabilities.ts; it must not hardcode a partial scope set in browser copy. Capability category order, labels, and descriptions must also come from the same manifest client; the settings panel may provide compatibility fallback rendering through that client, but it must not own a local category presentation table or /api/agent/capabilities fetcher. The panel's Pulse Intelligence surface summary is the same manifest projection: Pulse Intelligence Core, Patrol, Assistant, and MCP labels/descriptions plus surface affordance badges must flow through frontend-modern/src/api/agentCapabilities.ts from /api/agent/capabilities.surfaceContract, leaving the component to own only settings-shell layout. The visible external-adapter label in onboarding copy must also come from that manifest-derived capability posture instead of a hard-coded panel-local product label, so the settings shell cannot drift from the published agent surface contract. MCP capability-posture chips in that same panel must also flow through getAgentManifestSurfaceToolContract(manifest, AGENT_SURFACE_ID_PULSE_MCP) and getAgentSurfaceToolPosturePresentation, with the static inventory read only from /api/agent/capabilities.surfaceToolContracts; missing surfaceToolContracts entries must not make the browser infer MCP tools from raw capabilities. The panel must not own request / response capability filtering, call a per-surface projection alias, know the subscribe_events streaming exception, or maintain a local MCP tool count. The shared frontend manifest client must keep MCP onboarding on the generic surface resolver rather than exporting a Pulse-MCP-specific tool-contract helper. The panel's settings shell may arrange the external-agent setup hierarchy, but it must keep connected agents framed as optional access to Pulse context and Patrol work, with Patrol as the built-in operator that checks infrastructure, follows Patrol mode before acting, asks when approval is required, verifies outcomes, and records history, state that external agents use that same boundary, and make the canonical /settings/pulse-intelligence/agents#external-agent-setup route land on and briefly focus this panel with setup open rather than leaving the user at the generic API token inventory. Legacy /settings/security/api#external-agent-setup and /settings/security/api#pulse-mcp-setup links must remain accepted and may redirect to the canonical Pulse Intelligence External agents route. Normal API Access visits remain token-management first, and external-agent setup must not reorder the API token inventory because it no longer lives on that page. The Assistant settings default must keep setup mechanics behind a Show connector setup disclosure so it does not introduce a separate external-agent operator journey or make copied MCP config blocks or tool-contract proof badges the default visual weight of Pulse Intelligence settings. Its Developer details disclosure may show posture and policy summaries behind Patrol access model, but prompt, scope, failure-code, and tool inventories must sit behind a nested Live manifest details disclosure so the advanced surface remains navigable. The panel's manifest-owned pulse_operations_loop prompt row may expose the stable prompt id for client builders, but its visible label, description, and badge must keep the user-facing Patrol framing from the manifest instead of reintroducing operations-loop proof wording in the settings shell. The panel's explanatory onboarding copy must name published manifest-owned surface contracts as the publication boundary rather than suggesting raw backend capabilities become visible automatically. The visible token preset name in that setup must be Patrol external agent, not Pulse Intelligence agent; the latter may survive only as a route/model compatibility id. Manifest-backed stable failure-code summaries may use settings-shell chips or compact lists, but code selection and capability attribution stay owned by the API manifest client. Token setup handoff buttons or anchors in the Agent integrations panel are settings-shell chrome only: they may route to the API Access token creation section, but token preset semantics and required-scope derivation remain owned by the API/security boundary.
  5. frontend-modern/src/components/Settings/DataHandlingPanel.tsx shared with security-privacy: the data-handling settings surface is both a security/privacy trust surface and a canonical settings-shell presentation boundary.
  6. frontend-modern/src/components/Settings/dataHandlingPanelModel.ts shared with security-privacy: the data-handling settings model is both a security/privacy posture projection and a canonical settings-shell presentation boundary.
  7. frontend-modern/src/components/Settings/GeneralSettingsPanel.tsx shared with security-privacy: the general settings privacy panel is both a security/privacy control surface and a canonical settings-shell presentation boundary. The panel owns compact settings-shell framing for outbound usage telemetry, but its vocabulary must stay aligned with security-privacy: coarse deployment and lifecycle buckets, aggregate resource and outcome counts, coarse feature flags, and content-free Patrol, Assistant, and capability-API usage counters may be named, while hostnames, credentials, infrastructure identifiers, URLs, paths, locale, browser events, prompts, chat messages, command text, action output, token values, and personal information must stay explicitly excluded. Preview payload is the panel's primary action, because the exact runtime payload is the disclosure an operator can verify; the enable toggle and Reset ID stay secondary controls, and the summary copy opens with what the data is for before enumerating categories and exclusions. The shared settings shell no longer accepts a telemetryAction deep link that changes the preference on arrival; the preference changes only from the panel. The summary copy also states what the data is never used for (sold, shared, advertising, account or license linkage) in every locale, with the security-privacy disclosure as the source of those statements.
  8. frontend-modern/src/components/Settings/SecurityAuthPanel.tsx shared with security-privacy: the authentication settings surface is both a security/privacy control surface and a canonical settings-shell presentation boundary.
  9. frontend-modern/src/components/Settings/SecurityOverviewPanel.tsx shared with security-privacy: the security overview settings surface is both a security/privacy control surface and a canonical settings-shell presentation boundary. These settings panels consume the privileged security-status projection, while the shared status type also represents intentionally sparse public and authenticated tiers. Privileged posture booleans therefore remain optional at the transport boundary and must be normalized fail-closed before the settings shell derives posture summaries or hardening actions. The first-run shell must use generic host, Docker, and LXC bootstrap commands rather than probing public status for deployment identity.
  10. frontend-modern/src/features/platformPage/PlatformWindowedList.tsx shared with performance-and-scalability: the shared bounded list renderer is both a canonical platform-page primitive and a fleet-scale mounted-DOM performance boundary.
  11. frontend-modern/src/features/platformPage/PlatformWindowedRows.tsx shared with performance-and-scalability: the shared bounded table-row renderer is both a canonical platform-page primitive and a fleet-scale mounted-DOM performance boundary.
  12. frontend-modern/src/features/platformPage/usePlatformWindowedItems.ts shared with performance-and-scalability: the platform windowing controller is both a canonical frontend scroll primitive and a directional-runway performance hot path.
  13. frontend-modern/src/routing/routePreload.ts shared with performance-and-scalability, unified-resources: the app-shell route preload registry is a canonical frontend shell boundary, an authenticated hot-path performance boundary, and the entry point for the unified-resource Actions workspace.
  14. frontend-modern/src/stores/aiAvailability.ts shared with ai-runtime: the AI availability store is both the AI runtime session capability and the app-shell gate for Patrol entry points.
  15. frontend-modern/src/utils/platformSupportManifest.generated.ts shared with unified-resources: the generated platform support projection is both a canonical unified-resource platform union boundary and a shared frontend source/platform vocabulary boundary. It must expose the manifest surface_kind field so runtime lenses such as docker are not collapsed back into owning platform semantics. It must also preserve canonical projection lists from the governed manifest without page-local narrowing; for example, TrueNAS exposes both native vm, network-share, and app-container workloads through the same generated platform projection used by route helpers, badges, source filters, reportable-resource pickers, and type unions.
  16. frontend-modern/src/utils/sourcePlatforms.ts shared with unified-resources: the source platform normalizer is both a canonical unified-resource source adapter boundary and a shared frontend source/platform vocabulary boundary. That shared boundary must preserve availability as the agentless monitoring source for network-endpoint resources and settings presets, so source badges and platform/source type resolution do not fall back to generic when an endpoint is represented by ping, TCP, or HTTP probe data rather than an installed agent or provider API.

Extension Points

The planning-document status audit is a repository-governance support boundary owned here because it runs through the frontend lint entrypoint. It must derive its input set from Git-tracked docs/ files, skip separately governed subsystem contracts, and ignore untracked or ignored workspace notes. Local scratch documents cannot become mainline demand signals or block an otherwise valid push merely because their names end in _SPEC.md, _PLAN.md, or _CONTRACT.md.

Global Actions review uses the canonical shared Dialog, Button, Subtabs, Card, MetadataBadge, and mobile navigation primitives. The Open/History selector is an in-page sub-navigation surface and must compose Subtabs instead of recreating a segmented tablist in Actions.tsx; queue rows may own action state and resource semantics, but their frame and badge chrome stay on the shared primitives. Actions uses the full content width supplied by the app shell, matching Patrol instead of adding a page-local maximum-width container. The review's policy provenance uses a native keyboard-operable disclosure so the initial dialog layer stays calm without removing audit detail; intent, exact target identity, safety/authority posture, and fail-closed provenance warnings remain visible before expansion. The responsive route must preserve named tabs, keyboard focus, dialog focus containment, and phone-width overflow checks; journey 83 is the desktop/browser accessibility proof and is not mobile-device proof. The shared Dialog runtime in useDialogState.ts owns focus containment, body-scroll locking, background isolation, and focus restoration for every modal and drawer. While a dialog is open, every body-level surface outside the topmost dialog portal is inert, including surfaces added after opening; nested dialogs make their underlying dialog inert. The final close restores each surface's pre-existing inert state. Closing an overlay must restore its previous trigger with preventScroll; a plain focus() may make the app scroll shell jump to a virtualized or previously off-screen trigger just as the operator dismisses the overlay. Escape belongs to the top dialog and must stop the same keydown from reaching later global or shell listeners before it closes. Feature-owned dialogs may provide a stable fallback target when their original virtual row has unmounted, but they must use the same scroll-neutral focus contract rather than compensating with page-level scroll writes. InfrastructureWorkspace.tsx exercises that extension point after its shared Manage dialog closes: both the shared captured-trigger restoration and its delayed stable-row fallback must use preventScroll, preserve the app scroll offset, and leave focus on the originating Manage action at desktop and narrow viewports.

Assistant shell entry changes must keep Assistant contextual rather than generic: AppLayout.tsx and the command palette may expose a compact launcher, but that launcher must attach current-view context before opening the drawer and label the action around the current monitoring, Patrol, Alerts, or Settings view. The app shell may surface Patrol current-work pressure as a secondary count on the Patrol navigation tab, but it must not rename that tab to Needs Attention, create a Home-like action queue, or route the operator away from the Patrol route. Desktop and mobile accessible names should combine the stable Patrol label with concise open-work count context when a count is present.

SSO provider settings changes must preserve the shared Community-tier action path: SAML and OIDC provider creation stay on the same settings-shell control surface, while paid-plan copy and compatibility feature probes stay out of the frontend primitive boundary. The SSO provider settings shell is a fully migrated shared-action consumer: add, test, preview, copy, close, cancel, save, delete, edit, and dismiss controls must compose the shared Button, ActionIconButton, and CopyValueButton primitives rather than restoring panel-local <button> shells.

Feature surfaces under frontend-modern/src/features/ may own product-specific assessment semantics, but they must keep those semantics in their governed presentation helpers and render them inside the shared neutral Pulse surface language rather than introducing page-local verdict bands or nested cards. The Patrol operator home composes the shared PageHeader, Button, Toggle, MetadataBadge, native disclosure, dialog, tab, and neutral bordered-surface patterns. Patrol owns the meaning of background posture and the mutually exclusive Inbox, Protection, and Activity modes; the shared primitive boundary owns visible selection, Arrow/Home/End keyboard movement, roving tab focus, touch targets, responsive stacking, and visual consistency. Only the selected mode's panel is rendered. Plan-locked paid-mode discovery is not a daily-page primitive and must not be added beside the effective mode. Current API-owned decisions must follow the compact posture row in the default Inbox at every viewport. Objective configuration and verified/history surfaces belong to their explicit modes rather than following the queue on the same canvas. The compact Verified outcomes list consumes server-authored Patrol work receipts rather than filtering generic history in the browser. It may format the canonical capability name and verification time with shared presentation helpers, and it reuses the action resource presenter for durable resources that have left the live registry. It must preserve the server's verified-only membership, evidence class, newest-first order, empty state, and last-truthful-data behavior during a refresh error. Raw executor verification summaries stay in Actions history rather than becoming quiet-home copy. Selected Patrol decisions compose shared Button, ButtonLink, and CopyValueButton actions with native disclosures rather than introducing a detail-local control vocabulary. The primary reading order is summary, contextual Assistant explanation, affected resource, distinct impact, next step, and current lifecycle action. Raw canonical resource identifiers remain available through the shared copy affordance instead of occupying the reading surface. Evidence, protection, and timeline may share one collapsed disclosure when all typed detail remains reachable, keyboard-operable, and truthfully labelled. Compact actions retain the shared minimum touch target at phone widths even when their desktop presentation uses the xs or sm size. For Patrol, that includes the Open work description: it may use concise row-level guidance such as review evidence, approve a change, inspect automatic actions, or review verification results, but it must remain descriptive copy rather than another card, strip, or proof counter above the findings list. A findings list scoped to one resource keeps that description, the work-group chips and the filter counts on that resource: the description names it and counts its open findings, each count reads only its findings, the approval banner shows only their approvals, and the estate-wide check-health and coverage-stale chips stay in the unscoped queue, so nothing above the rows counts another resource's work. Watch-only forward-path guidance follows the same rule: the finding-row Switch to Ask first nudge and the Actions inbox Watch-only calm-state copy live in their governed presentation helpers (patrolControlPresentation.getPatrolWatchOnlyInvestigationNudge, actionPresentation.getActionsWatchOnlyEmptyState) and render as descriptive copy plus a single shared-primitive action inside the existing expanded-row and calm-state layouts rather than as new cards, banners, or proof strips. Feature-owned runtime hooks may also own non-visual side effects when those effects are part of the governed feature workflow. For Patrol, current-work action chrome must keep active findings in the Patrol findings workflow first; Assistant handoffs stay contextual actions on selected findings, approvals, or history records rather than the primary current-work CTA. When provider/model readiness blocks manual Patrol, the feature header must render the shared primary-action chrome as a Provider & Models setup link instead of a disabled run button that looks primary but cannot act. The Patrol hook owns the content-free pulse_operations_loop starter marker so render components do not fork telemetry or privacy behavior. When the shared operations-loop status projection reports contextual Assistant/external-agent collaboration inside the Assistant step, Assistant or Pulse MCP starter counts, Patrol control starter evidence, or Patrol control completed-loop or resolved-loop proof, feature presentation helpers may render those facts as compact title or step detail inside the existing journey layout. They must read primary patrolControl* fields first, then fall back to legacy patrolAutonomy* compatibility fields, and may fall back to legacy proActivation* fields only when the Patrol-control projections are absent. They must keep the neutral shared surface chrome stable and must not add page-local badges, nested cards, or alternate progress widgets for the same evidence. Feature-owned table drawers use shared disclosure and inline-detail primitives as local interaction state. Unless a surface has a separately governed deep-link write contract, opening or closing a row drawer must preserve the current document and URL instead of writing route state or reloading the page shell. Feature-owned sortable table headers must render real button controls inside the shared TableHead primitive and expose column state through aria-sort; the feature owner may define the sort keys and data comparator, but the header interaction must update that canonical owner state rather than forking table-local sort state or making header labels look clickable while inert. Storage pool headers are intentionally presentation-only rather than sortable headers: their one canonical sort interaction lives in StoragePageControls under View -> Order, so the table headings remain plain labels without a second hidden or competing sort affordance. Sortable header direction presentation belongs to frontend-modern/src/components/shared/tableSortPresentation.ts. Inactive sortable columns stay visually quiet; only the active column renders its ascending or descending marker. Workloads, backup, and shared platform tables must consume that helper rather than repeating dormant up/down icons in every header or defining page-local direction glyphs.

Shared filter/search primitives may provide the common shell, keyboard behavior, history, and reset mechanics, but the owning page or table must supply domain-specific visible copy, scope filters, status labels, and searchable field coverage. Platform pages must not surface generic "rows/resources" search affordances when the visible table is actually pods, VMs, datastores, apps, mail gateways, storage pools, backup jobs, or another product-owned object model. Add-filter controls that sit beside a page-level search box must prefer direct selectable filter values over a second nested search affordance; a page that needs filter-value search must keep that search inside the active filter chip or an explicit page-owned advanced selector. Platform-owned filter selectors must also exclude facet options from other platform scopes, even when the underlying shared surface is mounted from the same Workloads or Storage component. When every menu-backed filter is already active or has no selectable non-default value, the shared FilterBar must omit the exhausted Add filter control instead of leaving a disabled No filters affordance in the toolbar. Alert configuration tables follow the same primitive boundary: the alerts owner supplies platform-specific threshold groups and filter catalog values, while the shared FilterBar owns the chip, reset, and "+ Filter" interaction shape so thresholds do not reintroduce page-local search/tab chrome. The shared alert resource table's global-defaults row and card are the single editing surface for a section's defaults. A section that supplies globalDefaults must not also inject a synthetic resource row that mirrors the same record; the table renders defaults once, in both desktop table and narrow card layouts, and per-section metric columns must resolve to metric keys the shared column normalizer produces so the defaults editor reads and writes the same record keys the section persists. A global default that another setting decides while it is unset passes a globalDefaultFallbacks entry for that metric (label, title, whether the other setting is itself off, and the value switching on stages when it is). The TrueNAS Disks temperature uses it to read By type, following Disk temperature by type, with features/alerts/thresholds/trueNASDiskTemperature.ts giving each disk row its type's trigger from the unsaved editor state. resolveAlertResourceGlobalDefaultCell decides the cell for both layouts: unset follows the fallback, switching on from Off returns to unset unless the fallback is off, and an unset default with no fallback stays Off as the engine reads it. ResourceTable.tsx hands its props to the shared table state through getters, so the Custom badge and reset control follow defaults that load after the table mounts. Platform sub-routes that add native provider inventory must stay on the shared platform page and table primitives. The vSphere Networks surface routes through /vmware/networks, the shared platform tab model, the command palette navigation model, and the canonical table/detail primitives rather than a card deck or VMware-local page shell. Its rows are canonical network resources in the shared reportable/resource vocabulary, so source badges, resource pickers, command-palette search, table chrome, and detail disclosure must all consume shared primitives before VMware-specific presentation logic. A platform table's Health column, which says why a row's status dot is not green, renders through PlatformIssueReason (frontend-modern/src/features/platformPage/PlatformIssueReason.tsx): the first reason on the row's single line in the tone's colour, a count for the rest, every reason in the title and screen-reader text, and an empty cell for a healthy row. TrueNAS storage and the vSphere Datastores and Networks tables consume it, and the column stays desktop-only on each of them; a platform table must not hand-roll its own reason cell. On vSphere Datastores and Networks the Hosts list shows only from a 72rem container, where it reads instead of truncating, and the drawer keeps it at every width. vSphere tables that carry a vCenter column (hosts, health signals, activity) render it only when more than one vCenter is connected: the page passes singleVcenter (or the hosts table derives it from its scope), the header and cell drop together, and the inline detail row's column span follows. A table must not hide a column by filtered rows, because a search that narrows to one vCenter would otherwise make the column vanish. Patrol renders no primary assessment strip today. If one returns it is descriptive only; it must not render a Patrol-authored recommended next step, suggested prompt chips, or a secondary action band inside the assessment shell. Patrol-to-Assistant handoffs must carry only bounded evidence, resource references, and factual governed approval/action metadata as model-only context. Feature-owned Assistant handoffs may provide source context and safe metadata, but the shared drawer boundary must not turn those handoffs into frontend-authored prompts, tool routes, or remediation plans; the configured model owns tool choice and diagnostic reasoning after the request reaches the AI runtime. Such a compact Patrol assessment strip may include factual recent activity mix and trigger-mode labels when those values are derived from the Patrol run-history and status payloads. Those labels are summary context inside the same strip, not a replacement status card, CTA band, or page-local nested card.

  1. Add shared primitives in frontend-modern/src/components/shared/ Container update availability and executability are separate facts. When canonical action readiness refuses an update, the shared update control renders a passive amber Update available badge with a lock indicator, never an Update command button. Its hover, tap and keyboard-focus explanation preserves the server refusal and directs the operator to their container manager. In-flight actions retain their audited progress/review state. Filterable table surfaces that separate high-frequency narrowing from durable presentation preferences must compose frontend-modern/src/components/shared/FilterBar/ViewOptionsDisclosure.tsx. The shared disclosure owns trigger, expanded state, Escape focus-return, remembered-preference explanation, and responsive option-grid geometry; feature surfaces own the controls placed inside it. They must not restore a row of equally prominent preference toggles or a page-local popover shell. The disclosure renders in normal flow beneath the filter rows so durable settings push the affected table down instead of covering it. Its option grid uses compact, equal-width tracks that wrap from one column on phones to as many columns as the available desktop width supports. Floating, viewport-docked, and nested overlay implementations are forbidden for this persistent settings surface. Every option control fills its compact grid track, uses the shared control height, and divides segmented choices evenly; content-sized toggle islands with mismatched footprints are forbidden. An inline Columns picker expands as a full-width row below the compact controls and flows its checkboxes into responsive columns, so its option list becomes neither a narrow nested panel nor a tall single-column desktop list. Standard command buttons and button-styled route actions belong to the shared Button primitive family. Feature pages may choose the action label, icon, route, click handler, and contextual layout, but secondary/primary/ danger/outline/ghost button chrome, sizes, focus rings, disabled/loading behavior, and safe new-tab link behavior must come from Button, ButtonLink, and buttonModel.ts. Empty states, platform page notices, Patrol controls, settings panel actions, infrastructure setup controls, compact row actions, and other repeated command affordances must not copy local inline-flex ... rounded-md ... Tailwind shells just because the page needs a one-off action. The shared mdCompact size owns the common settings/action px-3 py-2 text-sm shape, the shared xs size owns the compact settings row-action px-2.5 py-1 text-xs shape, the shared settingsActionXs size owns compact settings/privacy px-3 py-2 text-xs action controls such as telemetry preview/reset buttons, and the shared iconMd size owns the settings dialog close-button h-9 w-9 icon shape. Positive completion or continuation actions such as infrastructure handoff and reporting exports use the shared success, successOutline, and successGhost Button variants instead of carrying page-local emerald action shells. Patrol approval and remediation controls use that same primitive family: Patrol owns approval/reapproval/denial/review/Assistant handoff behavior, while Button owns success, warning-solid, primary, secondary, ghost, disabled, focus, and compact action chrome. Shared error-boundary fallback actions are also command buttons: reset, reload, and retry controls must compose Button so emergency UI does not become a separate local button vocabulary. Update confirmation and progress modal actions are part of the same command boundary: cancel, start, retry, close, history, reload-now, and close-icon controls must compose Button or ActionIconButton instead of carrying modal-local blue, neutral, or icon-button class strings. Compact icon-only row, inline, and floating action controls belong to ActionIconButton. Feature surfaces may own the icon choice, click handler, label text, and layout slot, but icon-button size, tone, focus ring, disabled treatment, title fallback, and accessible name wiring must come from that shared primitive rather than page-local <button> plus inline SVG shells. Standalone machine row action triggers follow the same rule: the Machines table owns remove-agent semantics and menu placement, while ActionIconButton owns the compact muted trigger chrome. Patrol tool copy controls follow the copy-action boundary: ToolExecutionBlock owns the copied text and timer semantics, while CopyValueButton owns copied-state iconography, disabled handling, focus, and embedded-row propagation behavior. Global app-shell prompts are part of the same action boundary. frontend-modern/src/components/GitHubStarBanner.tsx may own its display timing, product copy, and GitHub destination, but its primary and dismiss controls must compose Button and ActionIconButton instead of carrying local floating-prompt button shells. The prompt must wait for the first return on a later day with connected infrastructure, consume at most one lifetime appearance, treat close as permanent dismissal, and never restore recurring snooze behavior. It must yield the browser session to an existing release, disclosure, or other low-priority app-shell notice and suppress itself for the remainder of a session that presents a blocking dialog. Low-priority app-shell consumers must coordinate through reserveLowPriorityNoticeSession rather than restoring component-local cooldowns that can produce consecutive prompts. Settings selection helpers such as ResourcePicker must use the same Button primitive for select-all, clear, and chip remove actions instead of restoring footer-local action shells. The picker must also keep initial all-resource hydration distinct from a genuinely empty estate: while the shared resource source is loading it presents a status placeholder, and only renders its empty-state copy after hydration completes. Intentional all-resource consumers use the shared websocket-first/canonical-REST race rather than waiting for every REST page before their first usable snapshot. Reporting surfaces must use the same primitive for retry, generate, and export actions rather than restoring large local CTA button shells. Self-hosted commercial plan, retry, activation, and clear-key actions follow the same shared Button boundary: commercial surfaces own the labels, entitlement state, and click handlers, while Button, ButtonLink, and UpgradeButtonLink own the primary, outline, warning, and upgrade/link chrome. Manual self-hosted key recovery is a secondary detail in that same boundary: settings surfaces may expose the fallback, but they must label it as license recovery/key recovery and keep normal checkout plus the Pro Patrol-mode setup path ahead of recovery mechanics or activation-key terminology. Hosted billing admin organization row actions follow the same boundary: cloud-paid surfaces own Suspend, Activate, Reload, tenant state, and mutation semantics, while Button owns the row-action chrome through the secondary sm and xs sizes. Security authentication settings actions follow the same boundary: security/privacy owns auth setup, password-change, credential-rotation, and read-only semantics, while Button owns the warning, primary, secondary, and settings-action chrome. Organization RBAC settings actions follow the same boundary: organization settings owns role creation, role editing/deletion, user-access assignment, feature-gate, and row-action semantics, while Button / ActionIconButton own primary, ghost, accent, danger, focus, disabled, and settings-action chrome. Organization overview, access, invitation, member, and sharing actions stay in that same primitive family: organization settings owns membership and share semantics, while Button owns primary, danger-outline, success-ghost, danger-ghost, disabled, focus, and row-action chrome. If a new surface needs a variant that the shared primitive does not expose, extend the primitive and registry guard rather than adding a page-local class string. Drawer header command and icon actions belong to that same shared Button primitive family. Workload and infrastructure drawers may own which actions appear and the action labels, but the h-8 Assistant, copy-context, close, and future drawer-header action chrome must compose DrawerHeaderActionGroup, DrawerHeaderActionButton, or DrawerHeaderIconButton instead of copying drawer-local button classes. The object-drawer subject row and its collapse interaction belong to ObjectDrawerHeader: the complete visible header is one semantic button with phone-safe target height, focus treatment, and a collapse chevron, while lifecycle or other object-specific action controls remain independent siblings above that button. Workload, infrastructure, shared inline, Ceph cluster, and Proxmox Mail Gateway detail consumers must compose this owner instead of leaving collapse on a small icon-only target or adding a local clickable wrapper around nested buttons. Copy-value affordances belong to the same shared button family. Feature surfaces may own the copied value, success/error notification, and adjacent product copy, but icon/chip copy controls must use CopyValueButton instead of recreating local copy icons, copied-state checks, or disabled empty-value handling. Compact information-card frames in drawer overviews, resource detail sections, web-interface URL editors, shared overview cards, and storage backup empty states belong to InfoCardFrame. Feature surfaces own the title, rows, actions, and sizing context, but the bordered bg-surface p-3 shadow-sm frame must be composed through InfoCardFrame, getInfoCardFrameClass, or INFO_CARD_FRAME_CLASS rather than copied as a page-local class string. Compact label/value facts inside those frames belong to InfoCardKeyValueRow. The shared row preserves endpoint alignment on phones, then switches to a fixed 7rem label track with the value immediately adjacent and left aligned on wider viewports. Consumers may opt into the sm transition when their own card grid already splits at that breakpoint; otherwise the shared lg transition is canonical. Feature surfaces own the labels, values, wrapping, and selection behavior, but must not restore drawer-width justify-between rows on desktop. This contract also covers secondary drawer facts in availability status and suggestion cards, resource change-history entries, Docker/PBS/PMG service support panels, Docker container-update management cards, action history, and the specialized Proxmox Mail Gateway drawer. Headers, status summaries, actions, disk capacity summaries, and RAID state pairs may retain intentional endpoint alignment; ordinary label/value facts in those surfaces may not. Read-only metadata chips belong to MetadataBadge and domain wrappers over it. Organization role and share-status chips must use OrganizationRoleBadge and OrganizationShareStatusBadge, so role/status tone mapping, pill shape, fit behavior, and whitespace handling do not drift across organization overview, access, and sharing surfaces. Patrol finding, investigation, approval risk/state, tool-call result, run-history/status-bar resource, outcome, snapshot, scoped-run, workspace-tab count, and contextual metadata chips must also compose MetadataBadge; Patrol remains the label/count/semantics owner, but the visible badge shell, sizing, tone vocabulary, and whitespace behavior stay in the shared primitive and shared-template-registry.json. Proxmox backup source/state chips follow the same boundary: storage/recovery owns backup-source labels and state semantics, while the visible chip shell and tone vocabulary route through MetadataBadge. Resource drawer identity chips (the IP address and alias chips and the alias overflow count in ResourceDetailSummary.tsx) compose MetadataBadge too. Badges stay on one line by default; an identity value that can outgrow a narrow drawer cell opts into the badge's wrap mode, which caps the chip at its container and breaks inside a token only where no space or hyphen fits. These values differ at their ends (...:pve1:113, a container hash) and a phone has no hover to show a title, so they wrap rather than truncate. A one-line alias used to overflow its fixed-width value cell at phone widths, where the cell's hidden overflow and the inline drawer's clipping content shell cut it off without an ellipsis. tests/integration/tests/04-mobile.spec.ts opens a Docker container drawer at phone width and checks that no element or text in its Identity section crosses either edge of those clipping boxes, apart from values that end in their own ellipsis. Workload backup freshness follows a separate shared presentation boundary: frontend-modern/src/utils/workloadGuestPresentation.ts owns the canonical tone and icon mapping consumed by workload rows and drawers. A recorded backup may be green or amber according to the configured freshness thresholds, but age alone must not produce a red failure treatment. The legacy never key means no completion time reported in current inventory: show cautionary Unknown, not a red no-backup verdict or an all-time absence claim. Unsupported workloads and templates remain neutral rather than being classified as unprotected. Inline detail content belongs to the shared detail-section primitive family. Feature surfaces may own the platform-specific rows, section labels, and source model, but section row shaping, empty-row compaction, value tone classes, table rendering, and inline close-action chrome must come from detailSectionModel.ts, DetailSectionTable, and InlineDetailPanel instead of local DetailField grids or provider-named reusable primitives. A compact detail row may carry optional bounded rich value content for links, tags, aliases, and address badges while retaining a canonical text value for titles, tests, and operator-readable fallback. It may also carry optional bounded progress metadata. A detail section may carry bounded full-width footer content for supporting evidence such as a compact change summary; that content remains inside the section card and must not be squeezed into the value column or detached into a feature-local card. The shared DetailSectionTable must render progress metadata through the CSP-safe ProgressBar while preserving the row's textual value as the primary operator-readable fact. Feature surfaces must omit the metadata when the measurement is unavailable rather than presenting an empty bar as 0%. Resource-detail drawer byte labels, integer labels, and count pluralization are part of that same primitive family: provider drawer models choose the fields and domain labels, but numeric detail values must route through formatDetailBytesValue, formatDetailIntegerValue, and formatDetailCountValue in detailSectionModel.ts. VMware vSphere drawer detail sections follow the same boundary as TrueNAS and Kubernetes: vSphere owns which rows are meaningful, while row shape, section shape, tone classes, and table rendering must stay on DetailSection, DetailRow, makeDetailRow, compactDetailRows, compactDetailSections, and DetailSectionTable rather than provider-local row/section aliases or custom vSphere card loops. Framed product table surfaces must consume the shared TableCard frame and TableCardHeader title/action band instead of composing page-local Card border, background, overflow, or table-title chrome. Feature owners may own the table data, filters, columns, and row behavior, but the outer product-table frame, section header band, and light/dark border treatment belong to frontend primitives so Infrastructure, Workloads, Storage, and Recovery do not drift visually. The shared Table primitive owns the horizontal scroll shell (overflow-x-auto plus touch scrolling), and the canonical .table-scroll-shell CSS explicitly suppresses vertical overflow so a phone gesture that starts over a table scrolls the page rather than an incidental nested vertical scroll range. Feature tables must not wrap it in page-local scroll containers just to restore table sides or mobile overflow. Headerless product tables, including alert history, still use TableCard for the outer frame instead of hand-coded rounded/bordered wrappers. Product tables already inside a canonical section frame, including storage pools, physical disks, and infrastructure settings source/configured-node tables, must use Table directly rather than nesting another card or scroll wrapper. If a framed table needs bounded vertical height, that constraint belongs on Table.wrapperClass so the shared table shell still owns overflow behavior. Resource-detail drawer tables that consume Table, including Docker Swarm services, Kubernetes namespaces/deployments, and PMG detail tables, inherit the same scroll-shell owner instead of carrying drawer-local overflow-x-auto wrappers. Other product table surfaces, including deploy wizard target tables, AI cost tables, Ceph tables, PMG resource panels, and PulseDataGrid, must follow the same rule: feature owners may pass wrapperClass for bounded height, border, radius, or scrollbar hiding, but they must not add raw table markup or local scroll wrappers around the shared table primitive. PulseDataGrid also owns its root frame variants: feature surfaces embedded directly inside an existing panel/card frame must use the shared frame="flush" mode rather than caller-local border overrides, horizontal-scroll wrappers, or negative margin compensation. Dense platform tables must also preserve useful operational context at phone widths. PlatformTableShell owns the responsive minimum-width policy, preserves any explicit base floor declared by the feature table, and otherwise applies the shared zero-width platform minimum before breakpoint-specific widths take over. Its class composition removes a consumer's redundant min-w-full token because the shared Table already fills the available width and that second minimum would override the phone-width policy. Platform consumers must explicitly select the three to five high-value phone columns that answer the row's operational questions, keep the identity column at the canonical 30 percent of the available width, use the shared compact label and column-width classes, and preserve the fuller tablet and desktop presentation. A sixth track is valid only when rendered values remain legible at the measured phone width. Duplicate state or severity text must not consume a separate phone column when the identity indicator already carries the same condition; the full value remains in the desktop table and inline detail. Consumers mark a normal-phone demotion with platform-table-phone-hidden on both its header and its body cell. The shared container rule owns the matching th/td visibility so the two cannot drift. A table that decides a column from its own measured layout instead drops the header and the cell under the same condition: the Docker Networks table had the class on the Subnets cell and the Driver header, so a phone showed bridge under SUBNETS. It now leaves both to its compact layout's Network details, and its browser proof counts visible headers against cells. Column visibility is container-led but md:w-[...] widths are viewport media queries, so between a 34rem container and a 768px viewport the phone-hidden columns return while no width applies, and every visible column took an equal share (a seven-column Kubernetes table left its name 83px). The shared shell keeps the 30 percent platform-table-name-column anchor in that band. Tables sized by a weighted colgroup keep their own widths there, since column widths win over header widths. A table whose columns need unequal room in that band (a long name beside short counts) declares each header's percent through platformTableBandWidthStyle or the sortable header's bandWidth, which sets --platform-table-band-width; the shell applies it only inside the band, after the name anchor, so md widths and phone widths are unchanged. A header whose full label outgrows its band column (Kubernetes Configuration's Lifecycle / trust) passes compactInBand to PlatformResponsiveTableLabel, which keeps the compact label it already shows on phones through the band; other headers switch at 34rem as before. A column whose content must read in full (a status badge) is sized for its widest label at the narrowest table in each range: a 34rem container for the band, and the table a 768px viewport leaves (about 695px) for md widths. The Docker Images Update check column needs 23 percent and 19 percent for that. The room comes from columns that lose nothing when narrower: a metric bar that never shows a sublabel or a collapsed actions menu, never a bar whose used/total sublabel would disappear. A weighted colgroup whose optional columns come and go is checked in every visible-column set, as the Docker container Updates column is in dockerContainerTableModel.test.ts. Below a 360-pixel content width, the shared narrow layout promotes identity to 40 percent and may remove one additional lowest-priority context column via platform-table-narrow-hidden rather than shrinking names and metric values into illegible fragments. That 40 percent is a floor on the name header, and a browser lays a fixed table out by the larger of a header's width and its column's, so a table cannot shrink identity below it from its colgroup (doing so only rescales every other column). A field that needs more room than the remainder gives it moves into the identity cell instead: Docker's container State words read on a second line under the name there, on every row so the rows keep one height, and CPU, memory and the Update control share the rest. The shared container rule then applies that stage across provider tables, direct Storage tables, and nested provider detail tables without a global reveal or page-local media-query exception. PlatformTableShell marks every owned table with the shared platform-table class; direct Workloads and Storage consumers and nested provider detail tables must apply that same marker so the phone contract cannot vary by rendering path. The marker owns a canonical 32-pixel summary-row rhythm across every provider and excludes inline detail rows, which remain content-sized. Summary cells must render one line only: secondary identity, raw provider labels, and descriptive context belong in an existing operational column, a supplemental tooltip, or its inline detail drawer rather than a stacked subtitle. The same holds for a note that explains a row's status dot: a TrueNAS service stopped while set to start at boot says should be running beside its state in the State cell, with the full sentence in the title and the drawer, not on a second line under the name. The Docker containers State cell likewise names what the dot means (Unhealthy, Starting, Exited (139)) in one word or two with the tone of the dot, instead of the raw running. Text-only operational rows may use getPlatformTableRowClass to declare that same shared rhythm explicitly; providers must not introduce local row-height exceptions. Mobile truncation may rely on a full-value row detail only when that row actually has a keyboard- and touch-operable disclosure; a non-expandable identity cell must otherwise expose its complete value without requiring hover. Default data rows stay single-line so scan density and row rhythm are not traded for automatic text wrapping. A summary value that must truncate sits in a block truncate span as the cell's only child, so it fills the fixed-layout column and ends in an ellipsis inside it on the row's text line. A capped inline-block max-w-* truncate span ignores the column width: it spills past a narrower cell with no ellipsis, wastes room in a wider one, and sits about 2 pixels above the row baseline. platformOverviewLayout.guardrails.test.ts holds the Docker, Kubernetes, Proxmox and vSphere table sources to that rule. An identity value whose distinguishing part sits at its end (a generated Kubernetes name, a registry-prefixed Docker image reference) instead splits into a truncating head and a kept tail (KubernetesNameText, DockerImageReferenceText), with the full value on the title, so rows that share a prefix stay distinguishable in a narrow cell. When a platform table row itself owns that keyboard- and touch-operable disclosure, SummaryRowActionButton removes the redundant visible chevron below the shared mobile breakpoint while preserving the button for screen readers and revealing it on keyboard focus. PlatformResourceDetailToggleButton applies that policy to provider-native platform tables, and workload rows opt in only while their compact row remains the disclosure target. Every disclosure-button consumer must declare that row-target relationship explicitly through the required hideWhenRowTappableOnMobile boolean, so a new table cannot silently restore a redundant mobile chevron. Rows whose click action does something other than expand details must pass false and keep a visible disclosure control. A grouped row may retain its summary focus or pin behavior while also owning disclosure, but when it does, the row click must perform both actions and suppress the redundant mobile chevron; summary interaction is not a reason for a one-off visual exception. They must not rely on a global rule that reveals every hidden column: phone priority remains source-specific because state, capacity, freshness, and recovery posture do not carry equal value for every resource type. Narrow viewports keep document-level overflow contained by Table and scroll the table itself. The shared scroll shell owns inline-size and paint containment plus horizontal overscroll containment so a readable table floor cannot widen or horizontally pan the application document. Feature tables must not squeeze every declared column into the viewport, override the shared floor, or add a second page-local scroll wrapper. Product-table subgroup/header rows must likewise consume the shared frontend-modern/src/components/shared/groupedTableRowPresentation.ts helper and .grouped-table-row CSS token contract instead of local bg-surface-alt or page-specific hover fills. This applies to grouped rows across Infrastructure, Workloads, Storage, Recovery, alert history, alert threshold tables, and Infrastructure Settings source-manager tables; feature owners may own group content and behavior, but not duplicate the subgroup band styling. Shared progress and metric-fill motion belongs to the frontend primitive CSS contract in frontend-modern/src/index.css. Generic progress bars must keep the CSP-safe ProgressBar / foreignObject shape and use the shared .progress-fill-frame and .progress-fill classes for width and color transitions instead of inline styles or page-local animation wrappers. The same global CSS owner must provide the prefers-reduced-motion disable path for these fills so feature surfaces inherit one accessibility policy. Numeric readout motion belongs to frontend-modern/src/components/shared/AnimatedNumber.tsx and its useAnimatedNumberState owner. Feature surfaces may opt metric labels and compact counters into that primitive, but must not create local counter timers, page-specific easing, or independent reduced-motion policy. The shared owner must also cap concurrent numeric animations and snap overflow readouts to their current target so a realtime estate update cannot make animation cost grow with the number of visible resources. Background colours come from the @theme inline surface tokens in frontend-modern/src/index.css and the bg-page utility beside them. bg-page paints the page backdrop (--color-bg-base) for a well that should read as the page behind a surface card: command and JSON blocks, the branding header preview, and the segmented filter track. A block that sits directly on a bg-surface-alt body, such as resource drawer content, uses bg-surface instead. No theme colour may be named base: it also colours the text-base font-size utility, and without it Tailwind 4 renders bg-base as nothing, which left those wells and the alert selection bar transparent. src/__tests__/App.architecture.test.ts fails if either comes back. Shared primitives must not reintroduce app-shell monitored-system capacity banners. Monitored-system grouping and ledger presentation belongs in the owned settings surfaces, while commercial plan explanation belongs in cloud-paid plan surfaces. Mobile navigation under the same shared boundary owns tab accessible names: icon components may keep their standalone labels, but the nav must treat those icons as decorative inside tab buttons so names come from the tab label plus meaningful badge counts, not duplicated icon titles. That mobile route surface must remain a bounded navigation rail rather than compressing every destination into unreadable tabs. It keeps the active priority platform destination together with Alerts, Actions, and Patrol in the fixed rail, groups the remaining platform and utility destinations (including Settings) behind a labeled More menu, and preserves active-state and badge context on that disclosure. The route rail is a semantic nav, not a tablist; the More menu must support first/last/active focus, arrow-key movement, Escape focus return, outside dismissal, and route-change closure. Shared grouped-resource presentation primitives must keep grouped resource labels operator-readable: count-led labels may aggregate repeated resources, but uncountable or category-like resource types such as storage must use resource wording instead of naive pluralization. Infrastructure table chrome on active platform/runtime pages must use mode-oriented labels for table presentation controls: grouped table mode is Grouped, not Cluster, because cluster remains a platform/resource concept for Proxmox, Kubernetes, and similar inventory details. The retired top-level infrastructure feature directory must not be recreated for table chrome ownership. Settings shell search copy belongs to frontend-modern/src/utils/settingsShellPresentation.ts. Shared settings search must not display non-actionable shortcut chips such as Any key; if the shell exposes a shortcut hint, it must name an actual key chord, otherwise the hint must remain unset so the shared SearchInput renders no shortcut chip. Native select state belongs to the shared frontend-modern/src/components/shared/FormSelect.tsx primitive. It must apply controlled value props after options are mounted so settings panels such as Discovery show the persisted option instead of falling back to the first option while the collapsed summary shows a different value. Its default select chrome (formSelect / formSelectCompact in Form.ts) sets appearance-none to match the text inputs, so it also carries the form-select-arrow utility, which draws the dropdown caret in the reserved right padding (and hands the native arrow back in forced-colors mode); without it every default FormSelect read as a text field. Callers that pass their own selectBaseClass keep their own arrow. The Assistant runtime controls in frontend-modern/src/components/Settings/AIRuntimeControlsSection.tsx are settings-shell chrome bound to the canonical useAISettingsState form and /api/settings/ai payload, not local browser state. Each must bind to a state.form.* field, round-trip through the field-by-field settings payload, and source its label, help, and summary copy from frontend-modern/src/utils/aiSettingsPresentation.ts rather than inlining strings or reaching for a bespoke fetch outside the canonical payload. There is no cloud-context-privacy control here: cloud context behavior is a fixed posture (see ai-runtime), not an operator setting.

  2. Add feature-specific presentation only when no shared primitive should own it. Feature surfaces under frontend-modern/src/features/ that display product labels must consume the owning subsystem's presentation utilities rather than hard-coding divergent page-local copy. Shared primitives and feature shells may compose those labels, but they must not become a second source of truth for alert, storage, recovery, infrastructure, workload, or adjacent product vocabulary. Table-mode segmented controls that expose a grouped/list view mode must use frontend-modern/src/components/shared/GroupedTableModeSegmentedControl.tsx so the shared primitive owns the Group by accessible label, Grouped and List visible labels, tooltip titles, and icons instead of each resource surface rebuilding that language with subtly different resource-specific concepts. Shared PageControls owns trailing filter-row actions such as toolbar display controls, utility buttons, Columns, and Reset. Controls that should wrap with the column/reset cluster must enter through toolbarTrailing instead of staying as loose filter-row children, and those controls must stay grouped when dense toolbars wrap so popovers remain viewport-safe instead of drifting off-screen from page-local flex behavior. The shared action rail must align to the trailing edge at wrapped desktop widths and remain separate from the filter-control wrap zone instead of waiting for a wide breakpoint, so Workloads, Storage, Infrastructure, and future dense toolbars do not strand Filter/Columns/Reset actions as an isolated second-row fragment. Shared FilterToolbarPanel owns default filter-popover geometry. The workload summary chart section that the shared chart show/hide action toggled is deleted and no other page exposes one, so the shared toolbar carries no chart show/hide action: the ChartVisibilityToggleButton primitive, its chart-visibility-toggle-button rule, and its chart-visibility-local-toggle-labels pattern guard are retired, and SharedPrimitives.guardrails.test.ts keeps them out. A summary section that returns through a governed product decision must register one shared show/hide primitive with an explicit pressed state rather than a page-local one-option segmented control or in-summary collapse chevron. Feature state hooks under frontend-modern/src/features/ own route-backed query state, selected item state, and data-window selection for their product surfaces; shared primitives and reusable presentation helpers may own viewport-safe chrome, focus treatment, pressed-state affordances, and accessible label builders for repeated controls. Estate-sized Proxmox backup tables follow that split: the storage/recovery feature owns its full-set-to-window projection and spacer placement while reusing the canonical useTableWindowing math; it must not invent a feature-local pagination or load-more button shell to avoid mounting the complete result set. Recovery timeline columns follow that split: storage/recovery owns the range, selected day, chart/table transport windows, and bucket data, while the shared frontend boundary owns the reusable button focus/selected styling and ARIA wording so columns expose singular/plural recovery-point labels plus selected state consistently. Frontend primitives must not fetch recovery data, infer recovery date ranges, or carry a parallel selected-day store just to render timeline columns. Compact, stable, mutually-exclusive filters with two to five options should use LabeledFilterToggleGroup as a responsive control: toggle buttons at wide desktop widths and the native select fallback below that. Dynamic and user/environment-sized option filters remain LabeledFilterSelect surfaces so estate-sized lists such as nodes never become button groups. A fixed six-state operational lifecycle may use LabeledFilterToggleGroup when every choice is a frequent triage action, the complete segmented group fits at its wide breakpoint, and the native select fallback owns all narrower layouts; this bounded exception must not be generalized to dynamic six-plus filters. Filters that change which other filters exist, such as Workloads Type, must stay in a stable primary filter band ahead of the dependent estate/data filters so changing the parent filter does not move its own click target or the adjacent primary filters; when multiple filters are expanded into button groups at wide desktop widths, each expanded group must have its own row rather than sitting immediately after another expanded group. User-facing filter options must use operator mental models rather than implementation categories: Workloads Type exposes a single Containers bucket while the system-container / app-container distinction remains an internal data/deep-link compatibility detail. PageControls owns the default stacked control deck for page-level filters: filter controls, display/chart controls, Columns, and Reset inherit one shared structured command deck with visible section boundaries instead of each page passing local controlDeckClass, action-rail, border, or background strings. Pages that have multiple semantic filter groups may set the shared filterControlsVariant="sectioned-children" mode and wrap those groups with pageControlsFilterSectionClass, but the deck chrome and trailing action section remain frontend-primitives owned. Those structured decks must give each semantic section a visible boundary so adjacent radio groups, scope filters, and display actions do not collapse into one hard-to-scan strip. Narrow consumers such as ColumnPicker must opt into their panel width through that primitive rather than layering competing width classes page by page.

  3. Add guardrail tests when a new shared pattern is introduced. Shared monitored-system primitives must prove they remain informational grouping or ledger surfaces rather than admission-freeze banners, cap summaries, or current / limit quota math. Shared modal scroll containment follows that same owner split. The dialog shell in frontend-modern/src/components/shared/dialogModel.ts must keep shared panels min-h-0, and page-owned modal bodies may use overflow-y-auto only under shrinkable flex columns instead of clipping lower fields behind a fixed-height shell. Shared filter popovers follow the same primitive-level ownership. The shared FilterToolbar panel class must render above nested card, table, and empty-state shells, and feature pages embedding those controls must make only their immediate control shell overflow-visible rather than forking local z-index or popover positioning rules. The shared navigation guide owns route-aware first focus: when it opens from a platform-owned product surface such as a recovery tab, the first highlighted step should match that route instead of always restarting at Dashboard.

  4. Keep shared infrastructure shell state on the reusable settings boundary: frontend-modern/src/components/Settings/useSettingsInfrastructurePanelProps.ts and frontend-modern/src/components/Settings/InfrastructureWorkspace.tsx must continue to derive provider counts and shared subtab copy from one infrastructure-settings source — via the unified aggregator through frontend-modern/src/components/Settings/useConnectionsLedger.ts — instead of creating provider-local summary fetches or VMware-only shell vocabulary. Phase 9 retired the old PlatformConnectionsWorkspace per-type shell; setup guidance should now use Add infrastructure plus source-strategy language for API-backed onboarding. The standalone connections-table presenter is retired; frontend-modern/src/components/Settings/InfrastructureSourceManager.tsx is the only landing-ledger presenter for configured infrastructure rows, and it must exclude agentless availability probes because those belong to frontend-modern/src/components/Settings/AvailabilitySettingsPanel.tsx. The first-run setup wizard inherits that same source-strategy vocabulary: step labels and completion copy must frame the final setup step as choosing the first infrastructure source, not installing a host. Successful token validation and security setup transitions should rely on the wizard progress state instead of transient success toasts that can cover the credential handoff. Generated first-run admin passwords must use browser cryptographic randomness rather than Math.random. That same shared shell boundary now owns the first-run posture for /settings/infrastructure: the landing route should read as one source-manager workspace with configured infrastructure instances first and no redundant monitored-systems ledger beneath it. The landing route may keep Add infrastructure in the Connected systems header while the ledger remains the first primary content. Discovery is optional setup after the ledger, not a competing first-viewport toolbar. Per-source add actions, including Install Pulse Agent, belong on the governed source rows, and Detect address stays inside the single API-platform probe path instead of a duplicate toolbar action. It may also show one compact posture line derived from the same unified connection rows so operators can confirm the top-level connected-system count, active state, actionable health, and limited coverage without opening a tour or second ledger. That line must expose the relevant install action when host telemetry is missing rather than expanding into a metric strip. Existing sources stay visible in stable source-catalog order, and add, detect, install, review, and manage flows open as secondary interactions from that same destination instead of taking over the whole page. The same source-manager workspace may show a compact fleet-governance strip and row-level fleet attention badges, but those badges must be presentation of the canonical /api/connections fleet object rather than another frontend-owned lifecycle classifier. Those secondary views must stay under the same single Infrastructure sidebar destination, but they may open in governed modal/dialog chrome when that preserves the persistent source-manager page behind them. That governed dialog chrome must also preserve inner form scrolling: InfrastructureWorkspace.tsx and ConnectionEditor.tsx keep the add/edit shell on min-h-0 flex columns so long credential forms scroll inside the modal body instead of trapping the lower fields below the fold. The same shared shell boundary now also owns grouped source-row composition. useConnectionsLedger.ts, InfrastructureSourceManager.tsx, and InfrastructureWorkspace.tsx must render attached collection methods as a compact labeled badge beside the owning system (API, Agent, or API + Agent), with the plain-language source phrase available through accessible metadata and fuller detail in the edit dialog, instead of duplicating the same machine across multiple peer groups or spending a dedicated Method column on implementation detail. The table-level product/system group rows in InfrastructureSourceManager.tsx must also use the shared grouped table row presentation helper, not local table-background classes, so source-manager grouping stays visually consistent with the product tables. That same shared shell boundary owns the landing taxonomy too: the primary grouping labels in the infrastructure manager must describe real platform/system owners, not collection methods. Agent-only machines belong in a standalone-host bucket, while Pulse Agent remains a collection- method label, install path, and detail-surface concept rather than a peer top-level pseudo-platform beside Proxmox, VMware, and TrueNAS. That same shared shell boundary also owns compact version visibility for agent-backed rows. The infrastructure source table must not grow a dedicated always-on version column for Pulse Agent; exact version text belongs in the edit/detail surfaces, while the landing table only surfaces a compact warning badge when an attached or standalone agent actually has an update available. That badge is the row's only version flag: the row adds no Version behind problem line beneath it, and the source band counts those rows as agent updates available rather than as needing attention. That same table boundary must reuse the System cell for compact standalone-agent identity such as Unraid 7.1.0; raw reported addresses belong in the governed Manage detail or an explicitly expanded cluster-member row, not an always-on diagnostics column. That same shared shell boundary now owns one canonical infrastructure destination in the Settings sidebar. InfrastructureWorkspace.tsx owns the source-manager landing inside that destination, while route-backed add flows and local edit flows stay single-purpose instead of stacking multiple page-level workspaces at once. The source-manager landing now also owns the explicit discovery strip for that destination. InfrastructureSourceManager.tsx exposes one optional Discover Proxmox systems status/action band after the systems ledger with scan state, saved scope, last result metadata, errors, Run discovery, Settings / Configure discovery, and candidate review when discovered sources are waiting. It must not start a network scan just because the page rendered. New-source admission belongs on the table's per-platform Add actions, the compact first-run/readiness actions, or the discovery band's explicit review action, and the direct address-probe utility may appear as first-run setup guidance instead of a second saved-network-scan command. That same shared shell boundary owns the cluster-member install action. InfrastructureSourceManager.tsx routes the cluster, member, and coverage-summary Install agent actions for a PVE cluster with a connected sibling agent into ClusterAgentDeployDialog.tsx, mounted once under a keyed <Show> so an install that outlives the dialog never reads a stale accessor, while every other row keeps the scoped installer hand-off. useConnectionsLedger.ts carries the Proxmox cluster name on cluster rows as clusterName, so the dialog keys the deploy API on the real cluster name rather than on a display label. Discovered API-backed candidates stay visible in the same platform-group table as configured sources, using the existing tree/table hierarchy instead of spawning a second discovery-only page or card stack. InfrastructureWorkspace.tsx must still open a new connection through frontend-modern/src/components/Settings/ConnectionEditor/ConnectionEditor.tsx, but the editor now serves as governed dialog content under the source manager rather than replacing the page inline. The ?add=pick route owns the search-first infrastructure source finder, ?add=detect owns the detect-from-address utility, and typed add routes jump straight into the matching credential slot through initialType. The picker must keep that first choice in recognizable system/service vocabulary, while the typed add dialog may use the shared source-strategy vocabulary after selection to explain API inventory, Agent telemetry, or API + Agent coverage. Agent-backed typed add routes keep the same governed dialog shell, but their embedded installer surface should stay focused on the selected system so the first visible command path does not re-expand into irrelevant platform choices. When an already-configured source is an agent-backed host profile, the source manager must group it under the operator-facing profile family and keep its add action on the same typed route instead of collapsing it back into generic standalone-agent copy. Credential slots are dispatched by the detected or manually-selected type and must still reach the canonical form body rather than diverging into a revived provider-specific workspace. For PVE, PBS, and PMG, the credential slot is frontend-modern/src/components/Settings/ConnectionEditor/CredentialSlots/NodeCredentialSlot.tsx and it must compose NodeModalBasicInfoSection, NodeModalAuthenticationSection, NodeModalMonitoringSection, and NodeModalStatusFooter inline under the editor shell rather than embedding the full Proxmox workspace (discovery card, configured nodes table, delete dialog, node modal stack). For TrueNAS and VMware the credential slots are frontend-modern/src/components/Settings/ConnectionEditor/CredentialSlots/TrueNASCredentialSlot.tsx and frontend-modern/src/components/Settings/ConnectionEditor/CredentialSlots/VMwareCredentialSlot.tsx and they must render only the connection form body inline under the editor shell — no connection list, no row actions, no surrounding panel chrome. Showing a ledger of other systems inside the credential slot is exactly the ledger-inside-editor drift this contract forbids. The configured-connections summary and source-manager rows themselves must render exclusively from the aggregator. InfrastructureWorkspace.tsx composes the platform-banded systems table from frontend-modern/src/components/Settings/useConnectionsLedger.ts (polling GET /api/connections). Table rows may open a governed edit dialog for mutable sources, but pause, resume, remove, last-error detail, and agent uninstall commands must live inside that owned edit surface or the shared row-action owner rather than returning to inline landing-page action clutter or a revived provider-specific detail page. When the backend marks a grouped Proxmox row with canonical cluster identity, the table primitive must render that cluster moniker as the row title instead of falling back to one sibling node hostname or reopening a standalone-host presentation for cluster-member agents. When that grouped row also carries backend-authored cluster members, the table primitive must render those nodes as child composition beneath the cluster row rather than flattening them back into peer top-level systems or hiding them entirely. The governed connection editor may assign an optional display name to each PVE cluster member only through its backend-authored immutable nodeIdentity. It must show the current native Proxmox name alongside an override, explain that clearing restores the native name, and keep the connection address as a separate control. Optimistic settings projection may update the display field only; it must not rewrite native name, host, IP override, credential, fingerprint, node identity, or provider node ID. Duplicate display values are valid presentation, so row keys, expansion, search ownership, edit targeting, and grouped-member joins must continue to use backend identity rather than label text. The same table shell must keep fulfilled rows visible across polling and manual reloads by using a retained-value query boundary, not app-level Suspense or a blank loading replacement, so configured infrastructure does not disappear while the next /api/connections request is in flight. The systems table and setup summary must count the same visible posture highlights they render, not hidden raw fleet signals. Passive attached-agent config or rollout handshakes whose only cause is a missing comparable applied configuration fingerprint may stay in the raw row model for deeper diagnostics, but they must not create duplicate cluster-parent badges or a Needs attention count when the visible row/member posture is otherwise healthy. That same landing-shell boundary also owns represented-host dedupe between the unified ledger and the discovery strip. InfrastructureWorkspace.tsx, frontend-modern/src/components/Settings/useConnectionsLedger.ts, and frontend-modern/src/components/Settings/infrastructureSettingsModel.ts must treat backend-authored hostname/IP aliases as canonical identity so an already-represented platform row, attached agent augmentation, or grouped member suppresses the matching discovered candidate instead of showing the same machine twice under hostname-versus-IP drift. Phase 9 retired the parallel reporting/inventory surface entirely: useInfrastructureReportingState, InfrastructureOperationsController, InfrastructureInventorySection, InfrastructureActiveRowDetails, InfrastructureIgnoredRowDetails, InfrastructureStopMonitoringDialog, and the per-type shells PlatformConnectionsWorkspace, ProxmoxSettingsPanel, ProxmoxDirectWorkspace, NodeModal.tsx, TrueNASSettingsPanel, and VMwareSettingsPanel no longer exist. The aggregator plus ConnectionEditor is the only path; no parallel reporting state, stop-surface dialog, ignored-row fallback, or per-type workspace may be reintroduced. connectionsTableModel.ts carries only the connection-scoped SystemManageAction variant — inventory-active / inventory-ignored manage kinds must not return. Active infrastructure settings and platform/runtime surfaces inherit that same source/platform vocabulary. Settings may label configured ingestion entries and endpoint probes as Source, while resource tables label their primary identity column as System. Lower-level unified-resource contracts preserve merged-source detail for tooltips, accessibility metadata, and routing. Collection methods such as Pulse Agent and runtime capabilities such as Docker may appear as option or detail labels, but they must not become the primary top-level system wording when a provider/API platform or reported host OS/appliance identity better explains what the operator is looking at.

  5. Keep settings deep-link route selection on the shared settings-navigation boundary. frontend-modern/src/components/Settings/settingsNavigationModel.ts and frontend-modern/src/components/Settings/useSettingsNavigation.ts must treat the canonical PBS and PMG Proxmox deep links as agent-selection authority even though those URLs resolve to the shared infrastructure-operations tab. Reloading or remounting on a PBS or PMG deep link must not silently fall back to the PVE selector state. Assistant OAuth callback compatibility queries such as ai_oauth_error and ai_oauth_success must route the bare settings root to Pulse Intelligence > Provider & Models while preserving the query long enough for useAISettingsState to consume and clear it, rather than normalizing the user back to Infrastructure and dropping the callback result. The canonical Agent Doctor browser route is /settings/infrastructure/agent-doctor with optional agents scope query parameters; it is the only live routed subpath under the /settings/infrastructure workspace path (all other infrastructure subpaths stay retired compatibility paths), it resolves to the infrastructure-systems tab, and the settings navigation hook must canonicalize pre-route agentDoctor=1 and agentUpdates=1 workspace queries onto that route with their agents scope preserved instead of rendering Agent Doctor as a dialog stacked over the workspace.

  6. Keep shared storage feature presenters on canonical platform truth. When reusable storage presenters under frontend-modern/src/features/storageBackups/ classify canonical resources for the shared storage route, API-backed virtualization datastores such as VMware must stay inventory-only datastores instead of inheriting PBS-specific backup-repository or protected-target copy from older fallback branches. Those reusable storage presenters must also keep primary issue copy separate from contextual impact copy. Composite posture fields may include dependent resource or protected workload impact, but shared table/presenter primitives must derive primary issue labels and summaries from explicit incidents, storage risk summaries, or storage-risk reasons so healthy rows do not render impact text as a warning. The shared resolveResourcePlatformType(resource) helper in frontend-modern/src/utils/sourcePlatforms.ts is the canonical reader for "what platform family does this unified resource belong to" and must be used by every frontend consumer that buckets unified resources by family (platform pages, filter resolvers, presentation pickers). The helper prefers resource.platformType when present and falls back to the resource's sources array via the existing source-platform normalization, so client-side family grouping behaves identically against mock fixtures and live backends that leave platformType empty on a subset of canonical resource types. Workload page membership must use the canonical platformScopes list when present instead of treating platformType as exclusive ownership. A Docker or Podman app-container can therefore carry both the container runtime lens and its owning platform page in routing/filter context, such as Proxmox when the runtime is detected inside a PVE LXC, while TrueNAS app containers stay scoped to TrueNAS even when their runtime metadata uses the shared Docker facet. That membership overlap is not permission to duplicate the detailed container table into every platform overview: Proxmox Overview keeps the default Workloads peer table to VMs and LXCs, and Docker-in-LXC evidence belongs as LXC drawer detail while /docker remains the canonical detailed Docker / Podman container lens. The overview table should not add peer rows, badges, or child rows for Docker containers; those signals compete with VM/LXC state and belong one click down. The default row may carry only a quiet icon/count cue beside the guest name to show nested runtime presence; names, metrics, state, and actions stay in the drawer or Docker lens. Shared platform-scoped storage presenters follow the same membership rule. A physical disk may be sourced only from the agent while canonically owned by a Proxmox node, so hidden provider-family filters must consult platformScopes before falling back to sources or platformType. Drawer-to-runtime navigation is still part of the shared platform-table affordance contract: when the LXC drawer exposes an Open Docker action for nested containers, that action must use the Docker host facet route state so the target Docker Overview opens scoped to the same runtime instead of a broad, visually unrelated container list. Primary navigation uses that same membership model: the Docker route is the container-runtime lens and may be labelled Containers in the shell, while shared source badges, filters, and runtime management copy continue to use Docker / Podman where the capability itself is being named. Kubernetes workload rows on /kubernetes/workloads must render through the Kubernetes-native workload tables rather than a generic infrastructure or inventory table. Pods render through frontend-modern/src/features/kubernetes/KubernetesPodsTable.tsx with Pod-native phase, readiness, restart, owner, node, image, and age columns; legacy /kubernetes/pods resolves to the same workflow. Controller rows render through frontend-modern/src/features/kubernetes/KubernetesControllersTable.tsx; legacy /kubernetes/controllers resolves to the same workflow. The table boundary preserves platform-native API fields for StatefulSets, DaemonSets, Jobs, and CronJobs, including targets, active/current counts, ready/succeeded counts, availability, exceptions, service names, schedules, and last run metadata. Kubernetes Overview keeps the cluster inventory above the native workload inventory so multi-cluster operators can orient and scope in one surface, matching the Docker Overview host-to-container flow. Selecting a cluster name writes the canonical cluster query parameter and scopes every Deployment, Pod, StatefulSet, DaemonSet, ReplicaSet, Job, CronJob, and autoscaling section beneath it; selecting the active cluster again clears that scope. The same URL-backed Cluster facet is available on Workloads, Services, and Configuration and composes with the existing namespace, search, and status parameters so bookmarks and shared links retain the complete scope. Changing cluster scope clears namespace scope to prevent a namespace from the previous cluster from silently hiding all rows. Cluster names remain separate from their external web-interface control and detail disclosure: the name scopes inventory, the adjacent external-link icon opens the persisted cluster URL, and the row continues to own drawer expansion.

  7. Keep shared source/platform vocabulary on the governed manifest boundary. frontend-modern/src/utils/platformSupportManifest.generated.ts must be the tracked frontend projection of docs/release-control/v6/internal/PLATFORM_SUPPORT_MANIFEST.json, frontend-modern/src/utils/platformSupportManifest.ts, frontend-modern/src/utils/sourcePlatforms.ts, and frontend-modern/src/utils/sourcePlatformOptions.ts must consume that generated projection instead of embedding divergent future-label lists, setup/onboarding path allowlists, host-profile labels, surface-kind guesses, readiness-state guesses, or presentation-only guesses, and frontend-modern/scripts/canonical-platform-audit.mjs must fail when the generated projection drifts from the governed manifest. The generated governance/readiness split is authoritative: supported platform arrays drive current support claims, while admitted platform arrays may keep route/navigation and add-flow vocabulary available without turning first-lab-ready entries such as VMware into supported-source copy. Kubernetes manifest projections must enumerate the native API-backed page sections the shared tab shell can expose, including controllers, networking, storage, config, policy, autoscaling, and events, so platform pages do not invent local support claims outside the governed JSON. The generated surface_kind is the machine-readable boundary between owning platform entries and runtime lenses: docker is a runtime-lens, not a platform, even when the container-runtime route stays available as a primary shell destination. The generic docker source-platform label is "Docker / Podman" in shared selectors, badges, and filter options so v5 Docker users can find the runtime surface while Podman-backed rows are not mislabeled as Docker-only; "Container runtime" remains the governed runtime family, not the primary customer-facing label. Identity colour is semantic, not page-local decoration: shared source/platform badges, host identity badges, and container runtime badges must use the shared presentation helpers so Docker remains on the Docker/Podman blue runtime tone, Podman uses its distinct runtime tone, Proxmox PVE remains orange, and those meanings do not drift across table rows, filters, drawers, or platform pages. Agent host-profile entries, including Unraid, stay in the generated agentHostProfiles projection and shared wrapper helpers; frontend primitives may render those labels for Pulse Agent install/identity copy but must not add them to the first-class platform union. The generated host-profile projection also carries runtime platform fallback metadata for shared explanation and parity with backend normalization, but frontend primitives must still render host-profile labels through explicit backend profile fields such as agentIdentity.hostProfile and unified-resource platformData.agent.hostProfile rather than prettifying presentation-only ids as platform values. Raw appliance identity aliases such as unraid-os belong only in the generated host-profile token list so shared helpers resolve them to unraid before presentation. Infrastructure System badges must append the platform runtime version when the payload proves that version belongs to the displayed platform identity, such as PVE pveVersion or a Pulse Agent report whose OS identity resolves to Unraid or Proxmox VE. They must omit the version rather than showing unrelated collector OS versions, such as Debian 12, beside an API-backed PVE badge. Surfaces that render Proxmox node versions, including the Proxmox page model at frontend-modern/src/features/proxmox/proxmoxPageModel.ts, must route raw PVE manager payloads through frontend-modern/src/utils/proxmoxVersion.ts rather than inlining page-local parsing or falling back to unrelated agent OS versions. System title metadata must apply the same identity rule: once the primary system badge names a platform with its version, source/method context may still add collection labels such as Pulse Agent, but it must not repeat the same platform again as an unversioned source badge. Proxmox Overview node rows also preserve provider availability explicitly. An offline or stale provider observation keeps the node identity, external link, detail affordance, and provider-scoped guest counts visible on desktop and mobile, while live uptime, temperature, bars, and sparklines render as unavailable. The row must expose visible Offline or Stale text rather than relying on an aria-hidden dot or opacity. Same-named cluster and member rows use Proxmox instance/platform scope for grouping, search ownership, and guest counts; a display label alone is not a cross-provider join key. Shared resource search must match the preferred display label, current and prior native Proxmox node names, and immutable node identity. Overview, storage, backups, alerts/history, infrastructure settings, API/websocket, and responsive/mobile rows must present the same preferred label while retaining native diagnostics; page-local formatting such as cluster (node) must not replace the backend-authored presentation. On the Proxmox overview, the committed Workloads search is shared state for both the guest table and the node table: guest matches retain their owning node, direct node matches retain that node, and unrelated nodes are hidden. The node table must use the same provider-scoped search vocabulary rather than receiving an unfiltered copy of the estate. Opaque unified-resource ids are not part of that visible search vocabulary and must not retain a node when the normalized guest table has no corresponding match.

  8. Keep summary interaction identity on one shared helper. Pages that expose row-hover or route-focus-driven summary emphasis, today the Workloads and Storage tables through useSummaryPageInteractionState in frontend-modern/src/components/shared/summaryTableFocus.ts, must derive page/entity scope through frontend-modern/src/components/shared/summaryCardInteraction.ts rather than letting each surface read hovered || focused from its own page-local ID source. No table previews or pins a group scope, and the shared helper and bridge carry none: Workloads grouped host rows are static identity dividers, and Storage pool-group headers are disclosure rows whose click opens or closes the group only. The Storage group pin went after the storage summary charts it was built to rescope were retired; the click that pinned an open group also collapsed it, hiding the member-row emphasis that was left. A grouped table that wants group hover or pin again needs a demand-ledger entry and must add that scope to this shared helper rather than reviving a page-local path. No summary card or chart strip renders today, and the bridge carries no chart-hover input; one that returns must consume that same resolved scope, and the chart under the pointer must enter its hovered series into resolveSummaryScopeState as a preview input that outranks row hover instead of keeping a chart-local hover island. Recovery is explicitly outside this interaction dialect: its retired posture-card strip must not return with row/group/chart hover behavior without a separate governed product decision.

  9. Keep any page summary page-scoped when table rows enter contextual focus. No infrastructure, workloads, or storage summary cards render today. If one returns, route-backed row selection may add a focused label and shared series emphasis, but the summary must continue to render the page-level series set instead of collapsing down to the selected row or replacing the global trend view with row-local empty states.

  10. Keep contextual row focus on the shared primitives. Same-route table drill-ins must reuse frontend-modern/src/components/shared/contextualFocus.ts for local scroll preservation, off-screen reveal, inline-detail lookup, and deliberate inline-detail reveal, and must resolve the active series through frontend-modern/src/components/shared/summaryCardInteraction.ts behind frontend-modern/src/components/shared/summaryTableFocus.ts, instead of rebuilding page-local Set filters, focused-label scans, drawer-aware scroll math, or ad hoc scroll restoration in each surface. frontend-modern/src/components/shared/summaryTableFocus.ts measures and scrolls rows only for deliberate focus of a row and its inline detail, and reveals a focused row once per focus change: its reveal effect subscribes to the focused series id and table root alone (a remounted table root starts a new reveal for the same focus) and runs the consumer's revealActiveSeries callback and the inline-detail scroll untracked, so page state the callback reads (live group lookups, expanded groups) cannot start another reveal, scroll the page back to the row, or reopen a group the operator collapsed. A reveal ends when its row and detail settle in view or its bounded retry window lapses, and filter changes do not re-run the consumer callback, so a row that stays filtered out past that window is not revealed when it returns. It must not keep window-level scroll or resize listeners or re-measure the hovered row on hover changes and scroll frames; an off-screen affordance for the active row needs a rendered consumer before that cost returns.

  11. Keep summary-linked table row emphasis on the shared primitive contract. Workloads, infrastructure, and storage rows that mirror the active summary entity must expose that state through data-summary-row-active and let the shared presentation in frontend-modern/src/index.css render the row emphasis, rather than carrying page-local sky or blue fill classes inside each row renderer. No table renders group-member emphasis: the data-summary-group-member-active rows and their index.css tokens went with the Storage group pin, so a group scope that returns must add its member emphasis to that shared presentation rather than a table-local outline, badge, or full-strength fill. Static grouped row headers on workloads, infrastructure, storage, recovery, and future grouped tables must use frontend-modern/src/components/shared/groupedTableRowPresentation.ts plus the .grouped-table-row CSS contract in frontend-modern/src/index.css, rather than rebuilding local bg-surface-alt variants with subtly different light/dark behavior or page-local left-accent markers. That shared grouped-table primitive owns the subgroup cell padding, typography, small metadata, and badge treatment as well as the row background token, so a future adjustment to the subgroup visual language changes every grouped product table from one owner. Inline table detail rows on platform, workload, and infrastructure tables must compose frontend-modern/src/components/shared/InlineDetailTableRow.tsx for the full-width row, surface-alt cell, detail padding, and row-click containment instead of rebuilding page-local TableRow / TableCell / div shells around each drawer. Storage-backed reusable row presenters under frontend-modern/src/features/storageBackups/ must also keep row height and alert accents on class/data-attribute presentation instead of runtime inline style maps, so the shared table contract stays CSP-safe on both steady-state and alert-highlighted routes.

  12. Keep retained-value data loading honest at the ownership boundary. Helpers that prevent a feature surface from falling through the app-level Suspense boundary during in-flight refresh should stay feature-local until multiple governed surfaces truly share the behavior. Once that boundary is shared, promote the helper into an explicit shared hook owner such as frontend-modern/src/hooks/createNonSuspendingQuery.ts rather than re-copying suspense-escape logic into each feature area or burying it inside one feature's private state model.

  13. Keep shared commercial warning banners truthful about destination intent. When a shared banner renders both explanatory and commercial CTAs, those labels must resolve to distinct owned destinations or section anchors instead of presenting two different labels that land on the same unscoped billing screen. Monitored-system capacity warning banners are retired; shared commercial banners must not render stale current/limit counts, paid-plan CTAs, usage summaries, or upgrade-impression telemetry from legacy monitored-system limit payloads. When a banner does need a review destination for a current paid feature, it must scope the operator into the usage-owned policy ledger rather than plan-selection intent or CTA copy that frames the flow as monitored-system-cap expansion.

  14. Keep AI availability bootstrap on the shared app-shell boundary. frontend-modern/src/useAppRuntimeState.ts, frontend-modern/src/App.tsx, and frontend-modern/src/stores/aiAvailability.ts must consume the backend-owned /api/security/status.sessionCapabilities.assistantEnabled fact instead of probing /api/settings/ai or /api/ai/* during ordinary route bootstrap. Non-AI settings panels may not initialize AI runtime state until an owned Patrol surface is actually open, and the app shell must not fork the AI-enabled fact across App.tsx, AppLayout.tsx, or page-level helpers. Non-critical app-shell prompts, including promotional or feedback prompts, must not use the shared blocking dialog stack because they must not look like required operational acknowledgement. AI-owned frontend surfaces that need shared settings or model-catalog truth must route those reads through frontend-modern/src/stores/aiRuntimeState.ts rather than each feature bootstrapping /api/settings/ai or /api/ai/models independently. Non-AI settings panels such as frontend-modern/src/components/Settings/useAgentProfilesPanelState.ts must stay on the app-shell AI availability fact instead of re-reading raw AI settings just to decide whether AI affordances should render.

  15. Keep Patrol shell composition and product-first provider vocabulary on the shared feature-presentation boundary. frontend-modern/src/features/patrol/PatrolIntelligenceWorkspace.tsx, frontend-modern/src/features/patrol/PatrolIntelligenceHeader.tsx, frontend-modern/src/features/patrol/PatrolIntelligenceBanners.tsx, frontend-modern/src/features/patrol/usePatrolIntelligenceState.ts, frontend-modern/src/features/patrol/patrolInvestigationContextModel.ts, frontend-modern/src/components/patrol/RunHistoryEntry.tsx, and frontend-modern/src/utils/patrolRuntimeActions.ts must keep Patrol assessment, verification, and findings primary; keep recent changes, learned correlations, and policy coverage out of Patrol-owned shells as backend or Assistant context, with no selected-finding or selected-run page section for them; and use Patrol/provider wording for the shared provider settings, provider model, and provider circuit-breaker affordances instead of generic AI labels inside Patrol-owned shells. The shared app shell in frontend-modern/src/App.tsx and frontend-modern/src/AppLayout.tsx must likewise expose /patrol as the canonical route and navigation target, while retired /ai browser entry points stay unregistered rather than a second Patrol-branded primary route. PatrolIntelligenceHeader.tsx must also keep the page heading's accessible name singular: when the PulsePatrolLogo appears beside visible Patrol heading text, it is decorative rather than a second label source. The Patrol workspace must not expose a generic Details/supporting-context panel for nearby activity, learned correlations, or policy buckets; those payloads stay backend and Assistant context rather than a default page section. Patrol initial data refresh failures must stay inside the Patrol feature shell as one compact stale-data retry banner; they must not replace the route with Suspense, blank loading, raw transport errors, or page-local diagnostic panels. If a Patrol Assistant handoff attaches recent-change records, the Patrol investigation-context owner must first normalize same-state records into changed-substate wording. The same shared feature-shell boundary owns the commercial-facing Patrol capability language: autonomy segmented controls and run-history/result labels must present the operator-facing policy levels as Watch only, Ask first, Safe auto-fix, and Autopilot, while legacy API names remain hidden from operators and compact controls do not collapse into unexplained shorthand. Patrol run-history rows must lead with what Patrol did or could not do before exposing trigger, token, tool-call, or raw trace details. Plan-locked Patrol controls must keep the free watch-only surface clean and must not render a Pro-absence explainer, a disabled paid-level matrix, compact Pro badges, or any paid-mode disclosure. The free Patrol working surface stays clear of paid-feature surfacing entirely; Pro discovery belongs in Settings, website/docs, and contextual at-need prompts, not beside the daily-use selector. The one allowed at-need prompt is a single finding-level Pulse Pro capability line in the expanded finding primary-action area for plan-locked installs on active critical or warning findings, with its upgrade action gated by the upgrade-prompt policy. Visible product copy calls the selector Patrol mode; compatibility route and wire identifiers may keep stable names such as patrol_control and patrolControl*. The Patrol header keeps the Patrol mode selector inside its collapsed Mode and automation disclosure and, for plan-locked installs, shows one inline Patrol mode line with the watch-only summary instead of a selector. The selector must stay on the selected mode and one plain summary, without a separate Limits disclosure or hard-limit matrix beside the picker. Shared feature shells must not invent their own Patrol safety thresholds, policy labels, or disabled-control explanations. The Patrol page header must consume the same effective control state, using watch-and-report copy for locked or Watch only mode and full governed-operations copy only for modes where that capability is actually available. Paid-control availability and commercial-plan copy must describe the same decision as choosing what Patrol may handle automatically; they must not ask users to decide how far Patrol can go or how much control Patrol has. The Open work workspace copy belongs to that same product-facing boundary: empty and descriptive text must explain what Patrol-found problems will appear there, what the selected control level allows, and the next useful operator action; it must not fall back to activation-loop, proof, queue, or verification-accounting language. Active Patrol issue rows may use shared definition-list and muted text primitives to show problem, affected resource, checked evidence, next step, and verification state inside the row, but that scaffold must not become a nested card, status strip, trust strip, or page-level proof block. A calm Patrol queue must not use shared compact-list or badge primitives to create protection-current, verification-waiting, schedule-freshness, drift, trust, or proof strips; empty work belongs to the plain empty-state and deliberate History affordance. When canonical Patrol evidence is stale, that same empty-state primitive may become warning-toned and direct the operator to run Patrol. The stale label may appear in the compact work-group row only alongside real current work; a calm queue must not repeat the same stale condition in both places. Monitor-context Patrol coverage posture must not use the shared compact list and badge primitives as a generic Proxmox overview or monitor-first launch-page proof strip. A future scoped monitor affordance may use these primitives only when it is attached to an operator action or selected Patrol context, uses distinct monitor labels, and does not become a nested card, generic dashboard strip, trust summary, or duplicate Patrol empty-work list. The Patrol page keeps the four-level control policy on its header and has no schedule and model drawer: schedule, trigger tuning, and the Patrol model override live on Settings > Pulse Intelligence > Patrol, and providers and the shared default model on Settings > Provider & Models. Those saves report through the shared notification toasts, not inline state: a rejected save raises an error toast carrying the backend's concrete reason rather than generic fallback copy, and a saved response whose Patrol readiness is not ready or degraded raises a warning toast naming the readiness summary, provider, and model instead of a success toast. On the Patrol page, an on/off save whose response reports Patrol not ready raises a fixed warning toast, and a failed Patrol mode save raises an error toast and restores the previous mode; neither renders inline save-failure state or offers an Assistant handoff. If a Patrol settings popover or inline save-failure state returns, it must stay viewport-bounded, expose an accessible dialog label, bind any model selector to the shared runtime settings and model catalog, show readiness context (provider, model, summary) next to a direct provider-settings action, and describe a saved but not ready response as a saved configuration issue rather than a failed save. When governed fixes are locked, the Patrol header presents Watch only as the effective mode with no enabled paid choice, and the Patrol state owner refuses any other level when a save starts. Its autonomy payload never carries full_mode_unlocked: PUT /api/ai/patrol/autonomy itself clears stale full-mode unlock and Autopilot activation state on every save below Autopilot, so disabled paid controls cannot leak stale permission into a save. The Patrol feature shell must also consume the Patrol-owned findings source for its findings tab, run-scoped findings panels, and tab badges so shared feature composition does not rebuild Patrol state by filtering the cross-product unified findings feed.

  16. Keep Pulse Intelligence settings product-first and page-scoped. frontend-modern/src/components/Settings/AISettings.tsx, frontend-modern/src/components/Settings/settingsHeaderMeta.ts, frontend-modern/src/components/Settings/settingsNavCatalog.ts, frontend-modern/src/components/Settings/settingsNavigationModel.ts, frontend-modern/src/components/Settings/settingsPanelRegistry.ts, frontend-modern/src/components/Settings/settingsPanelRegistryContext.tsx, frontend-modern/src/components/Settings/useAISettingsState.ts, and frontend-modern/src/utils/aiSettingsPresentation.ts must present that surface to operators under the Pulse Intelligence settings group as separate focused pages rather than as a generic AI Services shell or one oversized mixed form. Provider & Models owns API keys, default model selection, provider health/preflight, provider runtime budget/timeout, and usage/cost visibility. Patrol owns schedule, alert/anomaly triggers, runtime readiness, Patrol model override, and a simple Open Patrol handoff to the /patrol operator page; the actual watch/investigate/act/verify/record operator loop stays on /patrol. Assistant owns chat/tool permission, command-access, model override, and session maintenance. The AI Service Context page was retired with AI Discovery on 2026-10-11; its old route redirects to Assistant, and infrastructure discovery and onboarding stay under Infrastructure. The Provider & Models page must not carry a discovery summary, Patrol-control banner, or Patrol CTA; the Patrol-control handoff belongs on the Patrol settings page and /patrol, where copy describes Patrol autonomy in plain operator terms rather than exposing an internal operations policy concept. Settings-save feedback must preserve provider-specific preflight failures and successful save responses that carry Patrol readiness warnings, including the provider, selected Patrol model, failure cause, safe recommendation, and readiness summary when those fields are present. The settings shell may compose that safe backend diagnostic for display, but it must not infer provider remediation by parsing raw upstream error strings in the browser. Provider model identifiers, custom provider URLs, and API-key controls must compose the shared semantic form surface in both themes and must not be classified as website login credentials. The Provider & Models form and manual model identifiers opt out of credential autofill, provider secrets use the new-password autocomplete purpose, and the global form CSS preserves the active semantic surface and text tokens when a browser legitimately autofills another control. Dark mode must not expose the browser's pale credential fill or inject an admin username into a model identifier. Provider setup cards must describe provider families through the current backend-owned provider contract; DeepSeek setup copy is the V4 family and must not regress to old V3 or compatibility-alias wording. First-class provider cards on Provider & Models must remain model-driven through AI_PROVIDERS, AI_PROVIDER_CONFIGS, and the useAISettingsState provider payload mapping: adding direct chat-compatible providers such as Z.ai, Groq, Mistral, Cerebras, Together, or Fireworks extends those shared arrays/maps and the backend registry projection instead of introducing provider-specific JSX branches, local configured-state inference, or browser-owned default endpoint facts. The OpenAI card must present its API key as optional for a custom compatible endpoint and treat a saved base URL as configured provider state without inventing model-family prefixes in the browser. Readiness presentation consumes transport_healthy and patrol_capable: a reachable provider whose model did not emit Patrol tools is amber and explicitly remains usable for ordinary Assistant chat. Provider removal sends the complete remove_providers lifecycle mutation and rehydrates endpoint, credential, model, enabled state, and catalog from the response instead of clearing only the visible credential input. A blank Ollama keep-alive control means inherit the server default and must round-trip as blank. Local subscription-agent providers are the deliberate exception to credential inputs: their setup rows and first-run options render an explicit boolean opt-in, explain that Pulse uses an already authenticated same-machine CLI, and direct the operator to run provider readiness after saving. They must not ask for, accept, display, or imply storage of an OAuth token or API key, and must not present the opt-in itself as proof that the CLI login or selected model works. Their models remain normal provider-prefixed catalogue entries so Assistant, Patrol, service-context, and shared-default selectors do not invent alias parsing in the browser. The Ollama guided quickstart on that card (the copyable ollama pull command block, the hardware-expectation note, and the post-test next-step hint) renders the backend registry's suggested_model projection from the settings payload; the browser must not hardcode blessed model IDs or their equivalent tags, and the hint must compare the tested model against the server-authored suggestion set rather than a frontend literal. Provider connection controls are page-scoped: the global Pulse Intelligence enable toggle, provider readiness strip, and Test Connection action belong to Provider & Models; Patrol and Assistant subpages keep their focused settings plus reset/save actions without repeating provider health chrome. Those reset/save actions and save notifications must be page-scoped as well: saving Patrol or Assistant settings must not report Provider & Models settings saved or render a generic Save changes affordance that hides which Pulse Intelligence page owns the change. Runtime controls inside frontend-modern/src/components/Settings/AIRuntimeControlsSection.tsx must stay split by page-specific exports: provider runtime controls on Provider & Models and Assistant chat actions on Assistant. Assistant chat-action copy must name Patrol control as configured on the Patrol page, not as an Assistant command mode, because /patrol remains the operator surface for choosing how much autonomy Patrol has. Assistant-only controls such as execution permissions and session maintenance must stay explicitly labeled as Pulse Assistant controls, while Patrol schedule and trigger readiness live on the Patrol settings page and Patrol autonomy/control level lives on the /patrol operator page rather than drifting back into the provider shell. Session maintenance is limited to Pulse-owned conversation operations such as summarization; OpenCode-style file diff, revert, or unrevert actions must not appear in Settings unless Pulse owns a real governed infrastructure action-history/reversal contract for the affected resources. Shared/default model choice belongs on Provider & Models, while Assistant and Patrol model overrides belong on their respective settings pages instead of a generic advanced AI bucket. Each per-surface override must fall back to the shared default when left empty rather than silently using a hard-coded backend default, so Provider & Models stays the single place an operator picks the default model for both surfaces. Assistant model copy must describe chat, explanation, and review support; it must not present Assistant as the approved-fix executor because Patrol is the hands-on operator for checks, governed fixes, and verification. The shared shell must not show Pro-only autonomous execution as a default free-user control when upgrade prompts are suppressed; it may surface that option only when the entitlement is present, commercial prompts are explicitly allowed, or the current saved setting already uses autonomous mode and needs to remain visible for operator review. Provider model catalogs must remain curated on that same shell: frontend-modern/src/components/shared/AIModelPicker.tsx owns the searchable, notable-first model picker pattern, and AIModelSelectionSection.tsx must feed it configured-provider models plus the current manual selection instead of rendering raw provider catalogs as plain select options. The picker must also constrain its dropdown and internal result list to the available viewport height so settings model catalogs remain usable on mobile and tablet layouts with bottom navigation, and it must flip above its trigger when prompt/composer chrome leaves insufficient room below. Caller-owned alignment may choose left or right anchoring, but the shared picker still owns the fixed-position dropdown, viewport cap, search shell, result list sizing, and keyboard navigation model. Chat-owned selectors must reuse this shared picker instead of carrying a parallel dropdown implementation. Recent/priority model sections, external open-and-focus requests, selected older model visibility, route labels, explicit provider:model custom-route validation, and catalog-disclosure rows belong to the shared picker so Assistant, settings, and future model-selection surfaces do not drift apart. Unknown custom or recent routes may remain visible only when they have a valid non-empty provider and model segment; malformed route strings such as empty provider/model values, URL-shaped text, whitespace, or path-only payloads must be dropped instead of becoming selectable model routes. The shared picker must also mark the selected catalog, recent, override, custom, or inherited-default route as the current row with visible Current metadata and aria-selected; selected model state must not be communicated by background color alone. The model picker dropdown is a named search/listbox surface: opening it must focus search, the trigger must expose its owned listbox while expanded, and keyboard movement from search through the option rows must support current-row focus, filtered-result focus, catalog-disclosure focus, up/down, page, home/end, Enter/Space activation, and Escape return to the trigger so model choice and catalog expansion do not depend on mouse interaction. Picker-owned navigation keys, including Escape, must be consumed by the picker so parent shells do not also treat the same keypress as drawer or page-level Escape. Gateway-routed model choices must not look like direct-provider choices: the shared picker, System AI settings status strip, and per-surface inherited-default descriptions must render OpenRouter-hosted provider models with an explicit via OpenRouter route label while leaving direct DeepSeek/OpenAI/Anthropic/Gemini/Ollama selections unqualified. When a selected route also carries a shared-default or override badge, the shared picker owns that badge as separate metadata in both visible text and the button accessible name; labels must render as model via OpenRouter · default instead of fusing provider and badge text such as OpenRouterdefault. Platform-first top-level pages registered through frontend-modern/src/App.tsx route through the canonical app shell: each per-platform page surface owns its route, PlatformSectionTabs sub-tab navigation, and shared loading, empty, and error states, and its tabs are built around tables rather than estate overview or attention summary panels. Three tabs embed a cross-platform canonical surface with the platform scope locked: the Proxmox and vSphere Overview tabs embed WorkloadsSurface with forcedPlatform (the page owns the workloads state and the one WorkloadsFilter toolbar; the surface requires that state and takes only presentation props besides it, and the state reads its guests and nodes from the page's resource snapshot, with no enable flag), and the Proxmox Storage tab embeds StorageSurface with forcedSourceFilter. Those tabs must keep their guests, VMs, and storage rows on those embeds rather than a platform-local copy. Neither surface has an embedded or tableOnly mode, and no RecoverySurface exists: the aggregate Recovery surface was deleted on 2026-05-26. Every other tab, including Docker containers, Kubernetes pods, Machines, /truenas/storage (TrueNASStorageTopologyTable), the Proxmox Backups tab, and the TrueNAS Protection tab, renders a platform-owned table built from the shared platform-page primitives in frontend-modern/src/features/platformPage/sharedPlatformPage.tsx, even where a canonical surface could list the same resource types. Per-platform features must not fork their own table primitives, header layouts, or summary cards when a shared canonical surface already exists; new shared platform-page primitives live under frontend-modern/src/features/platformPage/ so the chrome stays reusable across families. Source-specific platform product surfaces under frontend-modern/src/features/, such as the Proxmox Backups tab, may own domain IA and row models in their product subsystem while consuming shared primitives for table shells, alignment, filter buttons, charts, and empty states. Frontend-primitives owns those reusable controls and guardrails, not the storage/recovery semantics that decide which PBS, PVE archive, snapshot, task, or workload-coverage rows are shown. Storage/recovery-owned Proxmox backup subcomponents may live under frontend-modern/src/features/ when they are listed in the subsystem registry and continue to compose the shared filter, table, chart, and empty-state primitives rather than local shell variants. frontend-modern/src/AppLayout.tsx may extend the PrimaryTab list with new platform or runtime-family entries, but primary navigation is a support-and-evidence-gated surface: rendered tabs, command/search destinations, keyboard shortcuts, and authenticated landing fallbacks must derive from the governed support manifest plus current runtime resource evidence. Supported platform/runtime families appear when evidence proves they are present; admitted-only, presentation-only, unsupported, or absent families stay hidden rather than rendering as disabled placeholders. The MOBILE_NAV_PLATFORM_PRIORITY ordering in frontend-modern/src/components/shared/mobileNavBarModel.ts mirrors that platform-first set only, so mobile and desktop navigation stay aligned without reintroducing aggregate Workloads / Storage / Recovery workspace tabs or the legacy Infrastructure entry. Frontend primitives owns the sole user-facing Machines IA contract for the support-manifest agent platform and agentless availability endpoints. The compatibility route, internal navigation id, and builders remain standalone / buildStandalonePath(); adjacent subsystem contracts may reference this owner for dependencies but must not restate the route, navigation, or landing semantics. Its primary tab, mobile priority, command-palette destination, and keyboard shortcut must all route through buildStandalonePath() and the PrimaryInfrastructureNavId standalone evidence gate; they must not create a generic Hosts, Nodes, Other, or mixed-systems bucket, and they must not include provider-owned platform nodes that are not canonical machine-page resources. The Machines page is a platform/runtime page, not a legacy Infrastructure page: it must use the shared platform tab, toolbar, table-card, and kind-aligned column primitives, and it must not reintroduce the old top-of-page InfrastructureSummary chart strip. The Machines surface must also remain secondary in the shell hierarchy when provider/runtime platform evidence exists: PRIMARY_INFRASTRUCTURE_NAV_IDS, desktop primary tabs, mobile primary priority, app-shell preload order, authenticated landing fallback, and command-palette ordering must prefer Proxmox, Docker, Kubernetes, TrueNAS, and vSphere ahead of Machines. The Machines surface may win those first/default positions only when the current estate has standalone Pulse Agent machines or agentless availability endpoints and no provider/runtime platform evidence. Provider/runtime pages must make narrower inventory scope explicit rather than presenting it as the complete machine estate: the Docker overview labels its runtime-specific table Docker hosts, and when those rows are backed by canonical Pulse Agent machines it exposes a View all machines action through buildStandalonePath(). Patrol workflow components under frontend-modern/src/features/patrol/ may compose shared Button and ButtonLink chrome for issue actions, but the workflow state, route anchors, single-finding direct-action selection, Assistant handoff, autonomy label, and multi-finding fallback semantics stay with patrol-intelligence; the canonical Patrol control anchor belongs on the visible selector, not the workspace shell. Setup-only readiness may hide run, schedule, model, trigger, and provider-repair controls, but it must not hide that selector or replace it with a setup/status explainer. Shared primitives must not grow Patrol-specific activation, autonomy, provider-settings, or Assistant-routing behavior. The default loop is a Patrol-owned watch / investigate / act under policy / verify / record loop. Active current-issue expansion is a task surface, not a history transcript: raw finding lifecycle rows may render in explicit all/resolved/history or selected-run review states, but not in the default active Patrol issue expansion. Compact Patrol status chrome may render work/health evidence passed by Patrol, but trigger/scheduling status remains header/control context and must not be repeated by shared default status primitives. Shared primitives must preserve that plain visible label instead of exposing internal assessment terminology on the default page. External-agent readiness from Pulse MCP may remain compact optional context derived from the shared manifest-client contract verdict and backend operations-loop externalAgentReady signal, but shared primitives must not create page-local MCP setup constants, token-scope checks, tool filters, readiness shortcuts, MCP readiness props, or a visible external-agent stage as the primary first-party loop. The journey's loaded progress state must come from the canonical operations-loop status projection exposed by the shared agent-capabilities frontend client, while shared primitives remain passive renderers of the state Patrol passes them. Patrol control starter, completed-loop, or resolved-loop evidence may change compact journey copy only after Patrol has derived it from that projection; shared primitives must not infer Patrol control or legacy Pro activation state from route anchors, billing state, generic Patrol recency, or MCP readiness alone. Shared presentation helpers may render the operations-loop state they receive, but they must not infer operations-loop progress from a generic Patrol run, recency timestamp, or MCP readiness alone; Patrol owns the issue-backed evidence model that decides when the loop can advance through Assistant, approval, rejected no-execution terminal decisions, approved-action verification, and external-agent parity.

  17. Keep user column sorting on platform tables on the shared sort fabric. Tables composed from frontend-modern/src/features/platformPage/sharedPlatformPage.tsx that offer user-facing column sorting must own it through createPlatformTableSortState plus PlatformSortableTableHead rather than page-local sort signals, ad hoc header buttons, or v5-style sort-drives-grouping designs. The shared fabric owns the interaction contract: click cycles a column's natural first direction, then the flipped direction, then back to the table's built-in order; sort state persists per table through usePersistentSignal storage keys; rows with missing values sink to the bottom regardless of direction; headers expose aria-sort and the arrow indicator consistent with the workloads table header through the active-only tableSortPresentation.ts helper; and header alignment stays on the canonical getPlatformTableHeadClassForKind helpers from columnAlignment.ts. A table's default order remains its page-model status-first compare until the user selects a column, and grouped modes (for example Docker's grouped-by-host containers view) sort within groups while grouping itself stays orthogonal to sort state. A saved sort applies only while its column's header shows: where a layout hides that column (Docker's Host and Engine while grouped, State under a phone row's name, Uptime on a phone), the rows keep the default order instead of reordering with no visible indicator, and the preference stays saved for the layouts that show it.

  18. Keep estate orientation inside the canonical controls operators already scan. Large-estate workload totals must flow from the unfiltered shared workload inventory into WorkloadsFilter.tsx, through the canonical FilterBar / FilterButtonGroup option-count contract, so each number sits directly beside the type or status label it describes without another fetch or a competing summary panel. Provider topology belongs in the existing table header; Proxmox derives cluster and standalone-node context through platformEstateOverviewModel.ts and supplies it to ProxmoxNodesTable.tsx rather than creating a page-level summary. A single-cluster estate shows that cluster's name in the same header slot in place of the count, kept visible while searching, because the repeated Cluster column is dropped. The existing platformEstateOverviewVisible preference now governs these inline totals from the shared View menu, preserving the global browser-persisted choice across platform workload surfaces. A page with adjacent totals, such as Proxmox's Nodes header, must own one visibility signal and pass it to both the shared filter and table instead of creating independently persisted signals that only synchronize after a reload. A provider host or node table must render before the workload filter, and the workload filter must sit immediately before the WorkloadsSurface it controls. It must never precede a provider table whose rows it does not filter; Proxmox and VMware vSphere use the same host-or-node, workload-controls, workload-table reading order. When a platform names the embedded workload collection, that title and its inventory count must enter WorkloadsSurface through the tableTitle slot. WorkloadsTable renders the slot inside the shared TableCardHeader, including the filtered-empty table state; platform pages must not leave the same title floating above the filter card or recreate table-header chrome outside the canonical Workloads frame. Large provider inventories use the shared createPlatformTablePreview and PlatformTablePreviewFooter boundary to keep the controlled workload table in the initial reading flow: Proxmox shows twelve node rows by default on larger layouts and six on narrow, phone and mobile layouts. The accessible, reversible show-all control belongs below the bounded rows, where a shared fade, remaining-row hint, and directional chevron communicate that the list continues; it must not compete with topology context in the Nodes header. Expansion is deliberate session state and must not persist a page-burying expanded default. Kiosk mode is the one exception: createPlatformTablePreview reads the shared kiosk signal, shows every row and renders no show-all control, because a wall display has nobody to press it. Docker / Podman, Kubernetes, TrueNAS, VMware vSphere, and Standalone Machines must continue using their shared PlatformTableToolbar counters and table headers; none may add a parallel estate card grid or provider-only spotlight surface.

Forbidden Paths

  1. Reinventing table/filter/toggle primitives when a shared version exists
  2. Feature-local styling forks of canonical shared components without explicit justification
  3. Direct imports that bypass shared presentation helpers where guardrails exist
  4. Top-level settings panels introducing bespoke page-level headers or outer framing instead of the canonical settings shell and SettingsPanel contract
  5. User-facing diagnostics or settings panels rendering maintainer/admin analytics such as commercial funnel, sales funnel, pricing/checkout conversion, or infrastructure onboarding telemetry. Those signals belong in admin-owned metrics surfaces, not the product diagnostics UI or customer frontend event emission.
  6. User-facing diagnostics panels rendering the native Pulse Assistant runtime as an MCP connection. Settings diagnostics must consume assistantRuntimeConnected and label it as Assistant runtime availability; mcpConnected, mcpToolCount, and "MCP Connection" are forbidden on the first-party diagnostics surface.
  7. Settings route models normalizing retired aliases such as /settings/operations/*, /settings/integrations/api, /settings/system-pro, /settings/workloads/*, or nested /settings/infrastructure/* paths back into current settings panels. Retired aliases must fail route eligibility instead of being kept as compatibility redirects.
  8. Platform pages implementing local estate metric cards, operational spotlight panels, or separate visibility preferences instead of projecting totals into the canonical filter and table-header controls.

Completion Obligations

Coverage-table polling must preserve the DOM identity of an unchanged logical row, including its focused expansion control, rather than keying rendering by replacement snapshot object identity. The isolated Chromium polling check in scripts/check-backup-browser-polling.mjs exercises this boundary with the production table, router and styles; it does not qualify full-app scrolling.

  1. Update guardrail tests when new shared primitives are added, including new Settings controls that drive backend verification surfaces (for example the Verify Patrol button in AIModelSelectionSection.tsx, which must drive the typed runPatrolPreflight client through useAISettingsState.ts rather than inlining fetch calls in the section component, must hydrate its result panel from the patrol_preflight snapshot on /api/settings/ai so the "last verified" state survives page reloads without forcing a re-click, must pass the form's pending patrolModel as the model override so the click tests the operator's unsaved dropdown selection rather than whatever was previously saved, and must surface a stale-cache warning when the form's selection differs from the cached result's model so the green badge cannot silently mislead). The readiness banner's tone and headline are pure functions exported from AIModelSelectionSection.tsx so this presentation is provable without mounting the settings shell, and a result that was never assessed — status: not_assessed, or the interrupted cause left by an operator cancel or a severed request — must render in a neutral "check did not complete" treatment. It must not use the failure treatment or the "model not verified" headline, because a run that measured nothing is not a verdict on the model, and it must not claim verification from a max_verified_mode recorded before the interruption (#1640)
  2. Keep top-level settings surfaces routed through the canonical settings shell and maintain both frontend-modern/src/components/Settings/__tests__/settingsArchitecture.test.ts plus tests/integration/tests/15-settings-shell-consistency.spec.ts The shared Settings content column must remain explicitly width-constrained and at rest across the lg breakpoint. Do not attach a transform animation unconditionally to the responsive panel container: entering a narrow viewport must not reactivate motion that translates or clips the active panel. Responsive proof must cover a desktop-to-390-pixel resize and the direct Plans & Billing route as well as ordinary Settings navigation.
  3. Keep Settings loading placeholders on the shared SettingsLoadingSkeleton primitive and the settings-loading-skeleton-shell registry rule instead of local animate-pulse block templates. Pure Settings loading indicators that are spinners rather than skeleton placeholders must stay on LoadingSpinner and the loading-spinner-shell registry rule instead of local border-t-transparent or border-b-2 animate-spin shells.
  4. Update this contract when a new canonical UI pattern is adopted
  5. Remove local forks after the shared primitive is introduced
  6. Keep shared feature-level presenters on capability truth. When reusable presenters under frontend-modern/src/features/ explain why a control, chart, or detail surface is unavailable, they must describe the owned identity or capability gap instead of prescribing a provider-local install path that conflicts with API-backed platforms like TrueNAS.
  7. When a settings route header and a top-level settings shell describe the same commercial surface, keep them on the same shared presentation owner instead of allowing route metadata in settingsHeaderMeta.ts or labels in settingsNavCatalog.ts to drift into independent title or description copy, and keep adjacent settings-shell referrals such as InfrastructureWorkspace.tsx on that same shared owner instead of reintroducing local “go to Pulse Pro” variants. That same shared owner must keep self-hosted commercial settings coherent when deliberately reached: the direct route, page shell, and any navigation shown for paid or recovery context use the shared Plans & Billing label, and the owned plan shell must foreground the active plan name plus available capabilities before secondary billing or recovery detail so paid upgrades can confirm their entitlement immediately after activation without making default Community look like it is missing an activation key. Routine plan and capability-status copy must stay product-facing: describe what is available on the instance, point failed capability checks to refresh the plan or open recovery, and avoid raw entitlement-payload phrasing or activation language as the normal setup story.
  8. When settings surfaces need informational, warning, success, or danger callouts, compose frontend-modern/src/components/shared/CalloutCard.tsx and register the consumer in frontend-modern/scripts/shared-template-registry.json instead of adding feature-local colored panel shells. Compact settings notices must use the shared scale="compact" density, with the registry entry as proof (frontend-modern/src/components/shared/SharedPrimitives.guardrails.test.ts runs the audit). Connection-editor status, feature-disabled, delete-error, and probe-result notices are part of the same settings callout boundary: the editor and credential slots own the source-specific lifecycle or API meaning, while CalloutCard owns the warning/success/danger shell and compact density. Update confirmation and progress modal notices share that same primitive boundary. The update flow owns version, prerequisite, root-access, restart, and error copy; CalloutCard owns the info/warning/danger shell, spacing, dark-mode tone, and icon layout.
  9. Keep hosted settings-shell framing imports safe for bundle initialization. Self-hosted billing titles, descriptions, and referral copy used by settingsHeaderMeta.ts, settingsNavCatalog.ts, and adjacent settings shells must flow through frontend-modern/src/components/Settings/selfHostedBillingPresentation.ts instead of importing generic commercial presentation helpers directly into hosted settings route shells. Contextual settings feature gates must use capability-owned presentation helpers and neutral paid-plan copy. They must not reintroduce Pro feature badge titles, Pro-suffixed option labels, monitored-system limit claims, or browser-local commercial/onboarding metrics wrappers in SSO, audit, reporting, AI controls, agent profiles, or shared warning banners.
  10. Keep shared settings-shell AI control copy capability-scoped rather than upsell-scoped. AIRuntimeControlsSection.tsx may describe read-only, approval-required, and autonomous action posture, but option labels and helper text must avoid tier labels or broad "executes everything" wording; paid capability availability belongs to entitlement-backed visibility and lock state, not local select copy. Provider & Models settings copy must keep Patrol autonomy distinct from Assistant chat actions: Patrol's hands-on control level belongs on the Patrol page, while the shared settings shell may only describe whether Assistant chat can run eligible chat actions. The panel carries no per-guest allowlist or denylist: guests the Assistant must not touch are locked with Never auto-remediate in the resource's Operator overrides, and AIRuntimeControlsSection.tsx points there instead of storing a separate Protected guests list.
  11. Keep first-session dashboard empty-state copy on frontend-modern/src/utils/workloadEmptyStatePresentation.ts, and make infrastructure setup guidance name the canonical destination explicitly instead of falling back to generic settings CTA labels.
  12. Keep the live first-session wizard on the canonical three-step runtime shape in frontend-modern/src/components/SetupWizard/SetupWizard.tsx (Welcome, Security, then Install), and keep the step indicator plus completion CTA language aligned with the governed infrastructure install workspace instead of regressing to a route jump that leaves the next action implicit. Preview-only follow-up surfaces such as frontend-modern/src/components/SetupWizard/SetupCompletionPreview.tsx must stay deterministic and scenario-driven: they may not poll the live /api/state runtime or inherit whatever connected systems happen to exist on the current backend, and browser proof for /preview/setup-complete must select explicit preview scenarios instead of ambient runtime state. That determinism boundary also covers the setup completion Pro activation pointer: every preview scenario in frontend-modern/src/components/SetupWizard/setupCompletionPreviewScenarios.ts must carry an explicit proActivation boolean so preview rendering never falls through to the live license probe, and the pro-unlicensed scenario is the canonical browser proof for the pointer. The pointer's localized strings (setup.completion.proActivation.*) are part of the first-session monitoring journey catalog and must stay in FIRST_SESSION_MONITORING_MIGRATED_MESSAGE_KEYS with non-identical DE/ES translations.
  13. Keep AI settings setup UI backend-driven: frontend-modern/src/components/Settings/useAISettingsState.ts and frontend-modern/src/components/Settings/AISettingsDialogs.tsx may collect provider credentials or runtime URLs, but they must not bake vendor model IDs into setup payloads. The shared settings shell should let the backend resolve the effective BYOK model and then render that returned state rather than guessing a model in the modal. Patrol-page surfaces must not add standalone trust strips to shared header or workspace chrome; high-signal trust facts may feed the Patrol-owned assessment readout, but shared chrome primitives stay free of the FindingsTrustSummary shape so adding new trust signals goes through the contract first rather than per-shell branching. Patrol refresh controls stay on that same feature-owned shell boundary: frontend-modern/src/features/patrol/usePatrolIntelligenceState.ts must make the refresh affordance generation-aware and timeout-bounded, so a slow supporting intelligence read cannot permanently disable the stale-data Retry control while Patrol findings and status remain visible.
  14. Keep shared filter primitives coherent with source-owned option hydration. Active platform/runtime pages and Settings infrastructure surfaces must keep canonical options visible in shared filter controls even when current results do not contain that option, so provider- or endpoint-scoped handoffs do not flash back to generic host-only language.
  15. Keep the first welcome screen in frontend-modern/src/components/SetupWizard/steps/WelcomeStep.tsx explicit about operator context. The shell must explain that the bootstrap token only unlocks first-run setup, state where the command should run, and adapt command/help text to detected Docker or containerized deployments instead of assuming the operator already knows which host or container owns the Pulse install. Bootstrap validation must remain an explicit operator action rather than auto-submitting on a token-length heuristic, and it must be single-flight so one successful validation advances the wizard exactly once even when click and keyboard submission overlap.
  16. Keep the settings-shell infrastructure landing path aligned with that same first-session story. frontend-modern/src/components/Settings/settingsNavigationModel.ts must treat /settings and the infrastructure settings tab as the canonical path to the bare /settings/infrastructure, which renders the unified Connections table, not to a separate install subview or to reporting/ control. The first-session story is owned by that table's own empty state and the Add infrastructure entry point on it, not by a second landing route, so first-time operators and returning operators see one consistent infrastructure surface by default.
  17. Keep Infrastructure and Workloads onboarding copy on the shared presentation owner in frontend-modern/src/utils/workloadEmptyStatePresentation.ts. Both the infrastructure empty state and the Workloads no-resources state must route first-time operators into the canonical /settings/infrastructure?add=pick source picker, let operators choose by recognizable system/service names instead of collection-method taxonomy, and avoid falling back to either passive “nothing here yet” wording or the retired install-first / Platform connections split. Workloads routes that already have canonical unified-resource infrastructure sources but no workload inventory must use a distinct no-inventory presentation that points operators at credentials, permissions, and collection status in the canonical infrastructure workspace instead of reusing first-run onboarding copy. That handoff is conditional on the session being able to open the workspace: the no-inventory presentation takes the session's infrastructureRead capability as a parameter, and a session without it must render neither the /settings/infrastructure action nor copy naming that page, directing the operator to an administrator instead. The gate reads the destination's own capability — the same one settingsNavCatalog requires for the Infrastructure nav item — so the offer and the nav gate cannot drift apart, and the variant stays a parameter of the shared presentation owner rather than a branch at any call site, including the surface's own inline fallback. Sessions whose capabilities have not resolved keep the action, so an administrator never flickers through the restricted copy. Inventory-source health itself stays visible to every session that can read monitoring data; only the Settings handoff is gated.
  18. Keep cross-surface investigation handoffs on shared route ownership. Feature shells such as Alerts and Patrol may decide which governed destination chips to render, but canonical href, label, dedupe, and infrastructure-fallback truth must stay in frontend-modern/src/routing/resourceLinks.ts instead of freezing raw route strings or provider-local link builders inside feature panels. Patrol workflow handoffs follow the same rule: start/continue Patrol control links must compose the route-backed patrol_control helper from resourceLinks.ts, single-finding direct action links must use canonical finding-presentation destinations such as the Patrol provider-settings route, while patrol_autonomy and legacy Pro activation URLs remain parser aliases only and verified review links use the plain Patrol history anchor. The Patrol finding-to-Assistant handoff has one owner, frontend-modern/src/components/AI/FindingsPanel.tsx; the Patrol state hook keeps no second copy of it, and the Patrol workspace passes no handoff callback because nothing on the page reacts to a handoff. UI surfaces must not duplicate the patrolControlStarter query string or write Patrol control or legacy entry-point starter telemetry from local click handlers.
  19. Keep retired self-hosted hosted-model and trial acquisition surfaces out of normal v6 GA runtime. Shared shells and helper-driven badges may continue to parse legacy payload fields, but ordinary self-hosted Assistant, Patrol, and settings flows must present provider setup as BYOK/local/self-managed and must not surface hosted-model credits, in-app trial starts, or generic managed-model claims.
  20. Keep shared contextual focus canonical after adoption. Once a summary or table surface enters route-backed contextual focus, future additions must extend frontend-modern/src/components/shared/contextualFocus.ts and its guardrail tests rather than forking another helper for workload IDs, resource IDs, or scroll-preserving same-route selection.
  21. Keep shared infrastructure/resource selectors on the canonical agent-facet truth. Shared primitives and settings-facing selector helpers must treat top-level TrueNAS appliances as agent-facet infrastructure via shared helper ownership instead of reviving a direct resource.type === 'truenas' branch inside page shells, selectors, or reporting-resource type helpers.
  22. Keep shared feature-shell Patrol run fixtures on the canonical run-record contract. When frontend-modern/src/features/patrol/ consumes Patrol run history, the shared normalized record must preserve provider-backed counts such as truenas_checked instead of letting feature-local fixtures or fallback objects collapse API-backed TrueNAS systems back into generic agent-host presentation. That same shared route-shell boundary also owns header-composition audit. frontend-modern/scripts/header-audit.mjs, .github/workflows/release-dry-run.yml, and .github/workflows/create-release.yml must prove the same shared top-level page-header contract before publication. The audit may follow local imports when a route shell composes PageHeader through a nested surface, and settings coverage must stay limited to top-level registry panels rather than every helper *Panel.tsx file. The canonical Settings shell therefore owns the shared PageHeader for support tools, and the retired top-level /operations/* browser path must not regrow a route-local heading, tab strip, or page shell for diagnostics, reporting, or logs. Because the dashboard route is retired, that audit must also discover live top-level pages from src/pages/ and may not keep a hard required-header entry for frontend-modern/src/pages/Dashboard.tsx.
  23. Keep the authenticated app root aligned with that same first-session path. That same shared-primitive ownership now includes contextual row focus. frontend-modern/src/components/shared/contextualFocus.ts is the canonical owner for nearest-scrollable-ancestor preservation, off-screen element reveal, and inline-detail lookup and reveal, while page/group/entity scope resolution belongs to frontend-modern/src/components/shared/summaryCardInteraction.ts behind frontend-modern/src/components/shared/summaryTableFocus.ts. Workloads and Storage row focus and the Proxmox page's node-to-guests reveal must route through those helpers instead of maintaining page-local copies of the same hover/focus rules. frontend-modern/src/App.tsx must land authenticated / and /login handoffs through this subsystem's provider-first platform landing contract: the first visible provider/runtime platform wins, and the Machines surface is eligible only when the current estate has standalone Pulse Agent machines or agentless availability endpoints and no provider/runtime evidence. The retired Infrastructure aggregate route and nested settings infrastructure aliases are not compatibility commitments: first-time operator setup must enter through the canonical Settings → Infrastructure workspace and its query-backed add flow, while provider evidence still owns the operational landing surface. frontend-modern/src/components/Login.tsx is part of that same pre-authenticated and first-session shell ownership: auth-check, setup fallback, and submit-pending loading indicators must compose frontend-modern/src/components/shared/LoadingSpinner.tsx through the shared-template registry instead of recreating page-local spinner shells. The subsystem registry must keep Login.tsx covered by the first-session-runtime-and-preview proof policy so login loading affordances cannot drift from the shared Settings, Patrol, AI, and primitive spinner contract. The authenticated app shell's boot-time route preloads must be owned by frontend-modern/src/routing/routePreload.ts so top-level cold-tab readiness cannot drift from the route-module preloader. The delayed boot-time set is bounded to the lightweight global Actions review destination. Alerts, platform, Patrol, and Settings modules load from current route or interaction intent instead of compiling an unseen workspace behind the active phone route. Route-module preloads and chart-cache fetches are separate shell responsibilities: the shared route preload inventory must stay module-only, while chart payload warming must route through the route or interaction that renders the chart. frontend-modern/src/useAppRuntimeState.ts must not prewarm retired Infrastructure summary-chart caches or eager Workloads chart caches as a generic authenticated-shell side effect.
  24. Keep relay settings shell copy on the shared presentation owner in frontend-modern/src/utils/relayPresentation.ts. The route metadata in settingsHeaderMeta.ts and the leading SettingsPanel in RelaySettingsPanel.tsx must reuse the same description and availability copy instead of drifting into separate rollout or pairing wording. Relay availability copy must describe the Relay tier boundary as Relay and higher plans rather than collapsing Remote Access back into a Pro-only feature.
  25. Keep shared settings-shell legal and docs referrals on frontend-modern/src/utils/docsLinks.ts. Shared settings surfaces such as AIRuntimeControlsSection.tsx must not hardcode GitHub main doc URLs for privacy, security, proxy-auth, scope-reference, or Terms-of-Service links.
  26. Keep shared settings-shell telemetry transparency controls on the governed general settings panel. Preview/reset affordances for outbound usage telemetry must stay rendered inside frontend-modern/src/components/Settings/GeneralSettingsPanel.tsx instead of drifting into route-local modals, hidden dev tools, or shell chrome that operators would not naturally inspect.
  27. Keep the short telemetry/privacy summary copy on that same shared surface accurate to the governed privacy doc. If the trust boundary depends on a specific retention window or on “IP addresses are not stored” rather than “IPs are never seen,” the summary copy in GeneralSettingsPanel.tsx must state those facts plainly instead of reverting to a stronger but inaccurate shorthand. The one-time schema-v2 upgrade disclosure must compose the shared InlineNotice banner layout on the existing post-update communication boundary, use shared button and external-link primitives, and route its Preview and Disable actions to the anchored General settings telemetry control rather than duplicating privacy state in shell-local UI.
  28. Keep maintainer commercial-event controls out of customer settings. The shared general settings privacy panel may expose outbound usage telemetry controls, preview, and reset affordances, but it must not render local commercial handoff event toggles, PULSE_DISABLE_LOCAL_UPGRADE_METRICS, or other commercial-debug controls as normal customer-facing preferences.
  29. Keep shared storage-route feature presentation on neutral capability truth. Reusable mappers and presenters in frontend-modern/src/features/storageBackups/ must distinguish inventory datastores from backup repositories so VMware rows on the shared storage route stay canonical to the admitted phase-1 floor instead of reviving backup-target, protected-target, or recovery-local semantics on a shared page. Those presenters must also source ZFS pool health from the canonical details.zfsPool payload (meta-first storage.zfsPool, flat platformData.zfsPool fallback) when building pool detail and bar summaries, rather than re-deriving device-level health from risk-reason strings or presenting flattened pool-state scalars as the full report. Ceph dedup is part of the same shared-presenter truth: cluster-internal pool rows must be consolidated into their mounting storage rows through consolidateCephClusterPoolRecords in frontend-modern/src/features/storageBackups/cephRecordPresentation.ts (lifting worse health onto the survivor) before shared storage tables render, instead of each table double-listing the same Ceph storage with conflicting raw-pool versus mounted-capacity accounting. Storage row models built by frontend-modern/src/features/storageBackups/storagePoolRowPresentation.ts must only carry fields the row actually renders; per-row source-platform badges and other identical-on-every-row decorations belong in the row expansion, not in StoragePoolRowModel.
  30. Keep infrastructure settings-shell API alternatives on the shared shell contract. frontend-modern/src/components/Settings/InfrastructureWorkspace.tsx, frontend-modern/src/components/Settings/settingsHeaderMeta.ts, and frontend-modern/src/components/Settings/settingsNavigationModel.ts must present the unified add flow as the canonical API-backed entry for Proxmox, TrueNAS, VMware, and future provider integrations instead of reviving top-level Direct Proxmox wording or shell-local provider routes. Phase 9 retired the Platform connections nomenclature along with the shells that owned it — there is no PlatformConnectionsWorkspace and no per-type ProxmoxSettingsPanel / TrueNASSettingsPanel / VMwareSettingsPanel to route through; the provider is a field inside one ConnectionEditor, not a destination.
  31. Keep the infrastructure settings connection inventory on one shared source. frontend-modern/src/components/Settings/InfrastructureWorkspace.tsx composes rows exclusively from frontend-modern/src/components/Settings/useConnectionsLedger.ts, which polls GET /api/connections. Provider connection counts and availability must derive from that aggregator, not from a top-level ledger plus parallel provider-specific fetches. The retired PlatformConnectionsWorkspace / TrueNASSettingsPanel / VMwareSettingsPanel panels must not be reintroduced as a second fetch path.
  32. Keep alert-history feature composition on the current owned state contract. frontend-modern/src/features/alerts/tabs/HistoryTab.tsx must react to the shared alertData() history state instead of reviving deleted aliases. Unified-resource resolution reaches frontend-modern/src/features/alerts/AlertResourceIncidentsPanel.tsx through the history state rather than a prop chain: useAlertHistoryState already receives getResource and re-exposes it, so the panel resolves a display name wherever it is mounted. That replaced the previous tab-passes-the-resolver rule, which only worked while the panel was a page-level sibling of the tab; it now mounts inside the history row that opened it (see the alerts contract, #1687) and the tab no longer renders it at all. Neither route may create another page-local resource lookup or a provider-specific handoff layer.
  33. Keep the alert-thresholds containers surface on the canonical shared owner. alertOverridesModel.ts, useAlertOverridesState.ts, and useAlertsConfigurationState.ts must surface API-backed app-container parents such as TrueNAS as first-class Container Runtimes, while ThresholdsTab.tsx must bridge function-valued selectors into ThresholdsTable.tsx explicitly instead of relying on spread-based adapter props that can collapse functions on the live Solid surface. Docker-only controls in ThresholdsTableDockerTab.tsx must remain gated to real docker-host resources instead of leaking onto platform-managed runtimes. Threshold host selectors follow the same single-owner rule: canonical platformType wins over secondary discovery facets, so TrueNAS and vSphere agent resources remain in their platform threshold tabs even when they retain Proxmox evidence, and only Proxmox PVE-owned hosts enter Virtualization Hosts. A missing canonical owner may use the legacy Proxmox scope resolver as a compatibility fallback.
  34. Keep shared commercial upgrade navigation typed and destination-aware. Shared paywall shells and upgrade actions must route internal billing or cloud destinations through frontend-modern/src/utils/upgradeNavigation.ts and frontend-modern/src/components/shared/UpgradeLink.tsx instead of guessing from labels, hardcoding target="_blank", or calling window.open(...) from each feature surface. Inline upgrade links may use UpgradeLink; button-styled upgrade CTAs must use UpgradeButtonLink so width, tone, focus, route/new-tab behavior, and opener preservation stay on the shared ButtonLink primitive instead of page-local Tailwind anchors or commercial helper class strings.
  35. Keep same-shell platform/runtime route transitions on retained shared state. Active infrastructure consumers may show full-page loading only before the first compatible resource snapshot exists; once a fresh canonical snapshot is already present in the shared app shell, top-level platform/runtime tab switches must reuse that state boundary instead of flashing a transient page takeover between tabs.
  36. Keep self-hosted paid-service prompts opt-in at the shared shell layer. settingsNavCatalog.ts, settingsNavVisibility.ts, shared upgrade link primitives, trial banners, monitored-system warning banners, and Patrol lock helpers must honor presentationPolicy.hideUpgrade by hiding paid prompts by default on ordinary self-hosted installs. Direct activation/recovery routes may render their owned content, but sidebar discovery, trial CTAs, plan upsells, monitored-system limit pressure, feature upgrade links, and plan-lock Patrol banners must require hosted mode, explicit handoff, or active entitlement. A lock the operator opens is not a proactive prompt: plan-locked lines in Settings and the locked long-range history states (HistoryChartOverlay and the guest drawer history lock, both through frontend-modern/src/components/shared/HistoryLockPlanLink.tsx) link to Plans & Billing whenever presentationPolicy.hideCommercial is false, and stay text only under demo and white-label runtimes. Cloud interest links from self-hosted plan surfaces must hand off to Pulse Account/public Cloud ownership rather than route to an in-product Cloud trial/signup page.
  37. Follow the AI runtime contract for retired AI Discovery behavior. Persisted/manual web-interface metadata remains the only row-link source.
  38. Keep settings sidebar search able to find pages by the vocabulary users actually type, not only by rendered copy. SettingsNavItem.keywords in frontend-modern/src/components/Settings/settingsNavigationModel.ts is the canonical search-only alias channel, matched alongside labels and header descriptions in frontend-modern/src/components/Settings/useSettingsAccess.ts; keywords are never rendered and must not become a second copy surface. The Assistant nav item must keep the external-agent connector aliases (mcp, model context protocol, external agent, claude, opencode, connector, pulse-mcp) so the pulse-mcp setup hosted on that page stays reachable from search, and the Assistant header description must continue to name external agent (MCP) connectors across locales. Proof lives in frontend-modern/src/components/Settings/__tests__/settingsArchitecture.test.ts and the parameterized search cases in frontend-modern/src/components/Settings/__tests__/useSettingsAccess.test.tsx.

Attention workbench shell contract

The app shell and Patrol attention workbench share the same canonical summary for desktop and mobile accessible navigation counts. The label remains Patrol; its accessible description may add the active attention count. frontend-modern/src/features/patrol/PatrolAttentionWorkbench.tsx uses native buttons, visible focus, selected-state semantics, focus restoration, narrow viewport ordering, and reduced-motion-safe behavior. Detail deep links use frontend-modern/src/routing/resourceLinks.ts and remain stable across reload. The temporary-suppression reason and duration are labelled native controls owned by FormTextarea and FormSelect; Patrol must not recreate their label, focus, responsive touch-target, or controlled-value shells locally. The Lasting decisions section reuses the same ownership: its four decision triggers are shared Button primitives in a labelled list, the inline confirmation note or rule reason is a FormTextarea, and the alert-only guidance link is a ButtonLink. PatrolIntelligenceSurface.tsx passes the Patrol findings accessor into the workbench; the workbench does not fetch or poll findings itself. The alert-only guidance names Patrol's open findings on the same resource from that accessor, and its Browse Patrol findings for this resource trigger stays a shared Button, rendered only when that list is non-empty. Open resource in the attention detail is a ButtonLink whose destination comes from buildResourcePagePath; the workbench hides it when no platform page can show the subject instead of linking to a page that cannot. The feature shell keeps the independent attention workbench mounted when Patrol is off or its model needs setup, while retaining the existing setup task above it. Desktop and narrow browser checks must show the task and attention list together, then allow a selected decision to open and return without changing the disabled Patrol controls or overflowing the viewport. The selected-detail header (title, queue position, Back, and the phone Previous and Next issue controls) is sticky top-0 with an opaque bg-surface/95 background, and the Patrol decision inbox section clips its rounded corners with overflow-clip, never overflow-hidden. An overflow-hidden ancestor becomes the sticky scroll container, so on phones the header scrolled away with the page and took the queue controls with it. Below lg the header pins against .app-scroll-shell; at lg it pins inside the detail panel's own scroller. The class contract is pinned in frontend-modern/src/features/patrol/__tests__/PatrolAttentionWorkbench.test.tsx. The objective brief and optional-context fields in PatrolObjectivesPanel share the same FormTextarea ownership contract.

The default queue shows severity, lifecycle state, plain-language consequence, resource, the required review posture, and observation age. It sorts severity first, then items with a governed approval or existing action, then newest observation, while retaining the server-authored membership. Provider evidence, protection posture, lifecycle controls, and history belong in selected detail; legacy Patrol analytics belong in collapsed supporting context. Unavailable and partial states must use explicit copy rather than success styling. Selected detail keeps the typed facts with named fields and plain unavailable-state labels rather than raw token pairs. On narrow viewports, selecting a row moves and focuses that decision context without horizontal page overflow. The narrow detail header uses a visible Back to list button, while the compact close-icon button is wide-layout-only; either path clears the attention deep link and restores focus to the source row. When the shared action inbox reports pending governed work, the attention header may render a primary /actions handoff beside its secondary refresh action. That handoff displays the action count but does not combine it with the attention badge or list membership.

The attention shell uses one responsive master/detail primitive. Its daily briefing owns the operator headline, three compact current-work counters, and a single highest-priority start action. Desktop keeps the decision inbox visible beside either the recommended-first-decision preview or selected detail; mobile hides the inbox while detail is selected and restores it before returning focus to the source row. Empty preview space must contain the recommended decision and an explicit review action near the top of the pane rather than vertically centering content below the initial viewport. The Patrol enabled/mode/check toolbar stays compact and does not repeat the page-level value sentence. Once review starts, the same master/detail primitive exposes a live queue position plus previous/next controls at both breakpoints. Its order stays identity-stable for the selected session, successful acknowledge or suppression announces the remaining count and advances to the next current item, and the final settled item returns to the calm inbox. Lifecycle copy must explain that review removes one occurrence from today's inbox while suppression is bounded; generic backend verbs must not leave the user guessing whether the alert was resolved, hidden permanently, or merely recorded as seen. Both the position label and calm evaluation age are reactive text nodes so detail navigation and refresh cannot leave their first-render values frozen. The selected metadata keeps latest-observation age visible, action verification copy stays generic across capabilities, and the list itself is not a broad live region; only errors, calm/empty transitions, and explicit completion announcements receive scoped assistive announcement semantics.

The selected attention detail may compose the shared Actions review for an eligible backend-authored offer. The detail owns only the expected postcondition, explicit-review warning, verification summary, and one review trigger. ActionReviewDialog remains the sole approve/reject/run and durable outcome primitive. After a decision or execution refreshes and replaces the detail subtree, dialog close must resolve and focus the current action trigger, not a detached element reference. Browser proof covers desktop and 390-pixel mobile layouts, reduced motion, screen-reader names, exactly one run request, and focus restoration for both confirmed and contradicted verification.

  1. Keep the Patrol model-readiness panel on the canonical AI settings state and shared model-selection surface. The panel must render connectivity, tool protocol, context quality, latency, and per-mode suitability as separate accessible evidence, support request cancellation, hydrate the latest compatible snapshot, and label failed or stale runs as evaluations rather than verified results. Safe auto-fix and Autopilot remain visibly not_assessed until their governed canaries exist.

  2. Frontend consumers identify a storage resource by storage.topology, never by a ZFS vdev layout. TrueNAS pools arrive as type: 'storage', so topology === 'pool' is the only pool discriminator available to the page model, and layout strings belong in storage.vdevLayout. Regression coverage: the pool identity boundary cases in frontend-modern/src/features/truenas/__tests__/truenasPageModel.test.ts and shows the vdev layout as the storage kind while topology stays the pool discriminator in frontend-modern/src/components/Infrastructure/__tests__/resourceDetailDrawerTrueNASModel.test.ts.

  3. The diagnostics export sanitizer owns the redaction boundary for the whole diagnostics payload, not just the fields that existed when it was written. Any provider failure string reaching the export must be passed through the IP redactor, including nested probe results and state reasons. Proofs assert on the serialized bundle rather than on the sanitizer's source, because the failure mode is a payload field added later that the sanitizer never learned about. Regression coverage: redacts PBS probe failures and state reasons in the exported bundle in frontend-modern/src/components/Settings/__tests__/diagnosticsModel.test.ts and keeps every PBS diagnostic failure string inside the export redaction boundary in frontend-modern/src/components/Settings/__tests__/settingsArchitecture.test.ts.

  4. An org switch must leave every live query surface with data for the new org, not merely cleared of the old org's data. Clearing state is not sufficient in createNonSuspendingQuery, because the reset writes signals the source effect does not track, so a consumer with a constant source and no polling would sit empty until it remounted. The org-switch handler therefore refetches, matching the clear-and- refetch shape every other org-switch handler in the app already uses. Stale pre-switch responses are still discarded by request generation. Regression coverage: refetches a constant-source query after an org switch and does not repopulate the cache when an old-org request resolves late in frontend-modern/src/hooks/__tests__/createNonSuspendingQuery.test.tsx.

Current State

Configuration import is bounded before dispatch

The existing Settings recovery import reads at most a 2 MiB selected file, allowing space for the export envelope and formatting. A larger file is rejected before File.text(), credential prompting or any import request. After parsing the supported JSON envelope, JSON string or raw CLI data, the complete serialized request (ciphertext, password, JSON escaping and UTF-8 bytes) must not exceed 1 MiB, matching configapi.handleImportConfig's request reader. File size alone cannot establish that the request fits.

Local size failures remain in the active import dialog, say no import was sent, preserve the inputs and allow an explicit corrected submission. They neither reload nor claim an uncertain applied import. Requests actually dispatched retain the existing non-retrying, workspace-bound and uncertain-outcome behaviour. The file control associates visible size guidance through aria-describedby.

useBackupTransferFlow.test.ts covers unread oversized files, the local file boundary, exact/under/over request bytes, multibyte/escaped passwords and explicit correction. BackupTransferDialogs.test.tsx checks the consuming accessible guidance. Desktop/narrow browser proof exercises the production transfer flow and recovery dialog with safe local responses; it does not establish native recovery or change any server limit, credential permission, archive or guest operation.

Retained monitoring status remains reactive (11 October 2026)

The existing Machines status dot subscribes to the shared relative-time clock for the unchanged five-minute agent-report fallback, agreeing with the standalone classification helper without a new report, row timer or API read. Machines health chip counts and selected results now use that same classifier at the same clock instant rather than the canonical last-reported status. Healthy excludes warning/stale rows, Degraded includes them and Offline retains the existing danger precedence. Muted/unclassified rows stay in All only. The status dot tooltip and accessible name expose that same current label (Online, Stale or Offline), not an online tooltip on an amber retained row. The canonical status, stale flag and timestamps remain observations, not derived health. Search, route-owned selection and removal behaviour are unchanged. The shared FilterButtonGroup keys buttons by logical value rather than transient option-object identity. Count, label, disabled, icon and order updates keep the same focused button while its option survives; withdrawn keys retire normally. Numeric and string values stay distinct, and selection uses the current entry. FilterButtonGroup.test.tsx covers those replacement/reorder/removal boundaries; FilterToolbar.test.tsx covers the live segmented-control label projection and forwarded selection at the shared wrapper boundary; the Machines clock/selection control exercises the production toolbar on desktop and touch-phone browsers. Attention sort classifies once per resource per call, preserving priority and numeric-name ordering without keeping a cross-call status cache. Debug source summaries read current status/error/last-seen by section ID, not a record captured when a retained payload section first mounted. Disclosure, focus and payload remain owned by that section.

The Machines, standalone model and Debug relative-age regression controls, plus browser-tests/pr2600-monitoring.cjs on desktop and touch-phone engines, cover clock-only staleness and arriving status replacement. These are synthetic source/presentation checks, not collector, workload recovery, whole-fleet performance, installed or release acceptance. browser-tests/machines-health-filters.cjs additionally covers expanded phone filters, exact freshness boundary, Healthy/Degraded clicks, arriving/withdrawn reports and preserved URL/search. The Proxmox identity and TrueNAS error guide checks still render the complete shipped guides through the production renderer; their per-suite fixture reuses only identical source/path renders, returns isolated clones and invalidates changed input. No renderer, guide bytes, assertions or test deadlines are weakened.

Retired frontend ownership

AI Discovery retirement and its compatibility boundaries are owned by ai-runtime.md. Deleted Discovery primitives have no current UI obligations. LoadingSpinner owns the remaining drawer loading boundary.

Manual update freshness

The update panel has a dedicated proof route to updatesPresentation.test.ts, with exact-content browser verification still required for rendered changes.

Check Now propagates explicit freshness through the shared update store and API client. A manual request arriving during a background check waits for it and then performs its own fresh read rather than returning early. A successful response records completion time, while failed refreshes retain the last successful result and its age. The settings release date uses the shared presentation helper, which omits absent, invalid and legacy zero timestamps. Cached update responses remain valid without a release date. Browser qualification covers the current settings route, pending/retry states, keyboard checks and date presentation at desktop, intermediate and narrow widths.

Provider tabs use compact canonical evidence during route hydration

Evidence-gated provider tabs consume the unified-resource owner's source-scoped type facets while rendering only the active workflow inventory. The shell keeps requested direct routes available during first hydration, then uses the settled facet bundle as the navigation truth. A temporarily absent row payload is loading state, not proof that the provider or workflow is empty; provider empty states may render only after the active route query settles. Docker / Podman, Kubernetes, TrueNAS, and VMware continue to reuse the shared platform tabs, loading, error, and empty-state primitives rather than adding a provider-local skeleton or navigation model.

Patrol objectives reuse the shared dialog, button, badge, and resource picker contracts

The Patrol retained-objective surface composes the existing shared Dialog, Button, MetadataBadge, and ResourcePicker primitives. It does not add a Patrol-only overlay, selector, badge vocabulary, or focus model. The modal keeps the standard backdrop, Escape, focus trap/return, bounded viewport, scrolling, and responsive footer behavior while the feature owns only objective-specific copy and orchestration. The same shared metadata badge renders a healthy server-owned proxy observer as Useful signal only with warning tone. The row keeps the backend explanation that the signal does not directly measure the full objective, while only server-authored covered state may render Watching in background. This is a presentation of the canonical objective contract, not a frontend inference or a new badge primitive.

System member rows are source-type aware

The Infrastructure source manager's member composition primitives now label by owning system type instead of assuming Proxmox: buildMemberRow threads the system type so vSphere host members present the "vSphere host" subtitle (table mode carries it in the row tooltip, exactly like Proxmox cluster members), and the expand toggle counts "hosts" for vmware rows and "nodes" elsewhere. No new member primitive was introduced — vSphere composition reuses the same expandable member-row rendering, status presentation, and merge semantics Proxmox clusters already use.

AI availability in the app shell is derived from the sessionCapabilities.assistantEnabled security-status capability, and aiAvailabilityStore.refreshEnabledFromServer() is the canonical way to re-derive it mid-session. AI settings save paths (setup modal, enable toggle, Provider & Models save, provider removal) and organization switches must call it so Patrol entry points and alert Investigate appear or disappear without a full reload; no surface may flip aiAvailabilityStore.setEnabled from settings state directly.

Patrol fix_rejected presentation is owned by the Patrol/AI finding surfaces that render the governed-action loop, while the surrounding badge, button, loading, and icon composition still uses shared frontend primitives. Shared primitive code must not special-case rejected Patrol fixes outside the standard metadata badge, button, and lucide-icon contracts. Setup-only Patrol action chrome is also governed by the shared Patrol workspace composition contract: provider-blocked states may show only the setup task and direct Open Provider & Models action, must suppress the duplicate readiness banner, and must not expose run-history buttons as a competing primary action before Patrol can check infrastructure. Patrol page setup banners must stay at operator level: render Patrol readiness payloads as Patrol setup issue or Patrol setup warning, keep provider/model context visible, and keep preflight/tool-call diagnostic wording inside Provider & Models rather than the first-party Patrol header. Pulse Intelligence settings now keep Provider & Models focused on provider setup, default model selection, health, budget, usage, and provider checks; it must not reintroduce the Patrol-control banner or Open Patrol control CTA that belongs to the Patrol settings page and /patrol operator surface. The Patrol settings page may still hydrate the cached Patrol diagnostic snapshot, but the rendered model-check panel must summarize it as model readiness rather than exposing preflight/tool-call implementation wording. The canonical Provider & Models browser route is /settings/pulse-intelligence/provider; /settings/system-ai remains a routeable compatibility alias for old deep links, while new settings navigation, OAuth callback redirects, Assistant repair actions, and Patrol provider-repair CTAs must emit the Pulse Intelligence route.

Shared loading indicators are part of the active frontend primitive contract. LoadingSpinner owns pure loading and action-pending spinner shells for shared primitive internals such as Button, PulseDataGrid, and HistoryChartOverlay, as well as Login, Settings, Patrol, and AI finding surfaces; local animate-spin spinner shells in those consumers are governed by the shared-template registry rather than page-local discretion. Update progress status indicators are included in that loading boundary: progress-stage loading must compose LoadingSpinner rather than local spinner SVGs. FilterButtonGroup owns feature table view toggles as well as settings segmented selectors: page-specific labels and selected values stay in the owning feature model, but the visible segmented selector shell must come from the shared primitive rather than a local bordered button group.

AI settings provider fields are a governed frontend primitive, not a provider-local form fork. The shared provider configuration section must render provider-specific controls from aiSettingsModel.ts extraFields, including Ollama keep_alive and the Z.ai custom base URL override, so Assistant and Patrol keep one settings shape across labeling, help affordances, helper copy, and persistence binding. The shared AI model picker owns model route search and presentation for Assistant surfaces. External open requests may seed an initial search query, but filtering, current/default route badges, recent routes, and custom route selection must remain inside AIModelPicker; callers should not duplicate that logic in command handlers or feature-local model selectors. Optional provider management actions belong in the same picker header as model refresh so Assistant and settings surfaces can expose provider repair without forking the model-list shell; callers own the destination, while the shared picker owns the button placement, labeling, close behavior, and keyboard-safe dropdown state.

The Patrol alert-trigger severity selector on Settings > Pulse Intelligence > Patrol (frontend-modern/src/components/Settings/AISettings.tsx) is built on the shared FormSelect primitive (label-for/id wiring, selectBaseClass styling hook) rather than a hand-rolled <select>, so its labeling and disabled-state affordances stay consistent with the rest of the AI settings surface. It sits in that page's Triggers section under the Alert-triggered Patrols toggle, renders only while that toggle is on, and is saved with the rest of the page by Save Patrol settings. The Patrol page carries no trigger settings; its header's Settings link opens this page when setup is not blocking.

The Patrol trigger and notification toggles on that page (Alert-triggered Patrols, Anomaly-triggered Patrols, Container update risk, Finding notifications) name each Toggle primitive through an explicit ariaLabel such as Enable alert-triggered Patrols, not ariaLabelledBy. Each accessible name must contain its row's visible label, so a change to a row's visible copy (for example "Alert-Triggered Analysis" to "Container update risk") must update its ariaLabel in the same change rather than leave a divergent accessible name. AISettings.test.tsx finds these toggles by name and checks that each name contains its visible label.

Kubernetes RBAC inventory (Roles, ClusterRoles, RoleBindings, ClusterRoleBindings) is part of the existing Kubernetes platform-page Configuration tab, not a new sidebar entry or top-level route, and the reporting-resource-type mapping at frontend-modern/src/utils/reportingResourceTypes.ts folds all four RBAC kinds into the existing k8s transport token alongside ConfigMaps, Secrets, and ServiceAccounts. Configuration tab rendering keeps RBAC summary fields (rule count, role kind / role name, subject count, subject Kinds, aggregation labels) bounded — individual subject names and full PolicyRule contents stay outside the rendered surface, mirroring the agent and unified-resource contracts.

No platform page embeds a Recovery workspace. The aggregate Recovery surface, its protection/events selector, and useRecoverySurfaceState were deleted on 2026-05-26, and TrueNAS opens on its Overview tab. Recovery rows render in platform-owned tabs whose controls stay on the shared filter-toolbar primitives: the Proxmox Backups tab filters through the shared FilterBar, and TrueNASProtectionTable.tsx through PlatformTableToolbar, rather than page-local tab strips or nested cards.

Cross-jump chip strips on alert and Patrol surfaces were retired on 2026-05-16 alongside the platform-first migration. The buildResolvedResourceSurfaceLinks and buildResourceSurfaceLinksForResource helpers (and the per-surface builders for Infrastructure / Workloads / Storage / Recovery hrefs) were deleted from frontend-modern/src/routing/resourceLinks.ts; the alert resource-incidents panel and Patrol findings panel that consumed them now keep investigation in-place through their existing handoff buttons and inline actions. Future cross-surface drilldown chips must not reanimate the legacy helpers. An alert card's single resource link opens the owning platform's page. It resolves the platform through the shared utils/sourcePlatforms.ts resolver, maps it to a page through PRIMARY_PLATFORM_NAV_SCOPE_IDS in features/platformNavigation/platformNavigationModel.ts, and builds the href with the routing/resourceLinks.ts path builders. It must not freeze raw route strings or keep its own platform alias table, because those drifted before: agent alerts linked to a /machines route that never existed, and TrueNAS incidents linked to Proxmox. Its fallback for alerts without platform metadata reads the resource type before any resource-id prefix, substring or message text, because a Proxmox guest id carries user-chosen cluster and node names and a message names the guest; those guesses sent a VM on node docker-01 to the Docker page, one in a cluster named agent to Machines, and a guest named vmware-test to vSphere.

Command palette and keyboard shortcuts moved to platform-first on 2026-05-16, and top-level aggregate workspace routes were retired on 2026-05-25 (frontend-modern/src/components/shared/commandPaletteModel.ts, frontend-modern/src/components/shared/useCommandPaletteState.ts, frontend-modern/src/components/shared/KeyboardShortcutsModal.tsx, frontend-modern/src/hooks/useKeyboardShortcuts.ts, frontend-modern/src/routing/routePreload.ts, frontend-modern/src/routing/navigation.ts). The legacy nav-infrastructure palette entry and g i chord remain retired with the unregistered Infrastructure route. nav-workloads, nav-storage, nav-recovery, and the g w / g s aggregate chords are also retired rather than hidden as compatibility commands. Platform commands remain nav-proxmox, nav-docker, nav-kubernetes, nav-truenas, nav-vmware (chords g p / g d / g k / g n / g v) plus a dedicated nav-kubernetes-workloads entry that lands on /kubernetes/workloads. The route-module preload registry and getActiveTabForPath matcher must not recognize aggregate workspace URLs as owned shell destinations. New palette commands and shortcut chords must flow through the same shell owners; do not reintroduce hidden platform families or retired top-level aggregate routes by reanimating legacy paths. The shared route-state helpers follow the same boundary: the workload and storage helpers in frontend-modern/src/routing/resourceLinks.ts may build query strings for an already-owned platform/runtime route, but must not export pathname builders for /workloads, /storage, or /recovery. The module exports no recovery query serializer: the Proxmox Backups tab owns its query through PROXMOX_BACKUPS_QUERY_PARAMS, and the TrueNAS Protection tab keeps its search, status filter, and sort out of the URL.

The shared table chrome now allows TableCardHeader to expose a right-aligned action slot, currently used by the Workloads/Proxmox metric display control. That slot belongs to the table header band and must not reintroduce nested cards or page-local toolbar wrappers inside TableCard. Proxmox host grouping also extends the shared NodeGroupHeader row pattern, which owns the header/table shell boundary rather than platform pages copying their own card headers. frontend-modern/src/components/shared/NodeGroupHeader.tsx renders node identity only (status dot, linked name, cluster and agent badges); in table-row mode that identity sits in one cell spanning the table. It carries no per-column cell renderer and no inline version, temperature, or uptime facts: host metrics and versions belong to the platform page's host table, so a group row must not duplicate them. Mobile navigation now recognizes proxmox as a first-class platform tab in the shared priority model so app-shell ordering remains centralized.

ResourceOperatorStateSection.tsx on the resource detail drawer overview tab uses createNonSuspendingQuery to fetch /api/resources/{id}/operator-state so the drawer's parent Suspense boundary does not flicker the page-level "Loading view…" fallback while operator-set state is in flight. New self-fetching sections inside the drawer must follow the same pattern (or wrap in their own local Suspense) rather than relying on createResource, which propagates suspension to the closest ancestor.

The Patrol page header copy lives in a single canonical helper at frontend-modern/src/utils/patrolPagePresentation.ts. The page-title tooltip on PatrolIntelligenceHeader.tsx must read from PATROL_PAGE_TITLE_TOOLTIP exported alongside the description rather than carrying an inline copy, so hover and inline never drift apart on what Patrol actually owns: watching infrastructure, detecting issues, recording findings, and escalating into governed investigation/action only when the selected Patrol mode allows it. The same PatrolIntelligenceHeader.tsx shell renders no trust-at-a-glance summary under the page title (Patrol's contract forbids a second header trust line), and no page renders a trust counter row: state.patrolStatus()?.trust only feeds the state hook's derived counts (historical regressions and prior Patrol work). The recency line beside the header actions also renders coverage alongside time when the canonical getPatrolRecencyPresentation helper returns resourcesCheckedLabel from the latest completed run. Render code must gate on <Show when={recency().resourcesCheckedLabel}> (truthy) so zero-coverage runs do not surface a misleading coverage phrase. The label reads checked N resources for every completed run with a positive count (full, targeted, follow-up, or errored) and never says verified. The page renders no assessment shell today; if one returns, its coverage caveats must come from the same run-history facts so they do not contradict the coverage the header line shows. The same header row may surface Trigger status when getPatrolTriggerStatusSummary returns a runtime-relevant value from the Patrol status payload. That text is page-owned operational metadata inside the existing header row, not a new shared primitive, nested status card, or secondary verdict band.

Shared frontend empty states, thresholds empty states, and discovery guidance that mention the Infrastructure settings destination now consume frontend-modern/src/utils/infrastructureSettingsPresentation.ts for the canonical Settings → Infrastructure label and source-strategy copy. Shared primitives must not fork that string or revive removed nested route labels.

SettingsTab no longer includes infrastructure-connections or infrastructure-install. The single infrastructure-systems entry in settingsNavCatalog.ts, settingsPanelRegistry.ts, and settingsNavigationModel.ts replaces both. Panel routing within the infrastructure area uses InfrastructurePanelStep in-page state. The shared monitored-system warning banner has been retired. Ordinary hosted and self-hosted sessions must not render app-shell monitored-system capacity warnings, plan-review links, or upgrade-impression telemetry from stale finite policy data. Shared alert presentation surfaces (OverviewTab.tsx, HistoryTab.tsx, AlertOverviewActiveAlertsSection.tsx, AlertHistoryTableSection.tsx, AlertHistoryTableAlertRow.tsx, AlertOverviewAlertCard.tsx) no longer accept hasAIAlertsFeature or runtimeCapabilitiesLoading props. Feature gating for AI alerts flows through the shared entitlements layer; surfaces must not re-introduce per-surface capability fetch props. Recovery's retired posture-card strip remains outside the shared hover-synchronization dialect. Per the Extension Points constraint, Recovery must not introduce a new summary-card component with row/group/chart hover wiring without a separate governed product decision. frontend-modern/src/components/Storage/useStorageSummaryCharts.ts now owns the reusable polling/caching state for storage summary history, while frontend-modern/src/features/storageBackups/storageCapacityDeltaPresentation.ts keeps pool-growth label/tone formatting inside the shared feature presentation layer. The storage page must keep reusing those shared owners instead of rebuilding storage-history timers or byte-delta formatting inside row components. Storage row alert reasons stay in the same feature presentation layer: storageRowAlertPresentation.ts turns the row's open alert into a short headline (and a compact phone form), storagePoolsTablePresentation.ts carries it on the row model with its severity tone, and StoragePoolRow renders it in the State cell through the existing issue-text classes, with no inline style and no new row height. That same shared alerts feature boundary now also owns legacy shared-storage override migration. frontend-modern/src/features/alerts/alertOverridesModel.ts and frontend-modern/src/features/alerts/useAlertOverridesState.ts must canonicalize per-node shared-storage override keys such as Main-pve1-ceph-pool, hashed /api/resources storage ids, and Ceph pool storage rows onto the storage metrics target id before the thresholds table derives rows, so old Ceph override records and newly projected Ceph pool overrides survive the v6 feature-shell path instead of silently disappearing from the live editor. Docker container override identity follows the same single-owner rule. dockerContainerOverrideIdCandidates in frontend-modern/src/features/alerts/alertOverridesModel.ts is the only builder of container override keys (re-exported by thresholdsResourceModel.ts): it leads with the stable docker:{host}/{containerName} key that survives container recreates (#1601) and trails the legacy container-ID, short-ID, unified-hash, and slash-tail forms as lookup candidates. Threshold rows carry that chain as overrideIdCandidates/overrideStorageId so every mutation path (toggle, connectivity, offline state, edit, remove) strips the historical keys and writes only the name key; docker surfaces must not derive container override keys locally from resource-id parsing.

The frontend already has several guardrail tests. The next step is to keep turning repeated local patterns into explicit shared primitives with hard usage bounds, including provider-backed alert-history wording. frontend-modern/src/features/alerts/helpers.ts, frontend-modern/src/features/alerts/tabs/HistoryTab.tsx, and frontend-modern/src/features/alerts/OverviewTab.tsx must present VMware- backed host and VM incidents with the shared resource-incident vocabulary and existing alert-history shells instead of introducing VMware-only labels, badges, or panel copy just because the underlying signal came from vSphere. The same helper labels the host-offline alert type Agent Not Reporting: that alert only means an agent's reports lapsed, so drawers and the alerts page must not call it a host outage beside a node row that is still online. docker-host-offline keeps its own Host Offline label. That same shared settings and modal boundary now also owns the public usage-data vocabulary. frontend-modern/src/components/Settings/GeneralSettingsPanel.tsx, frontend-modern/src/components/Settings/useSystemSettingsState.ts, and frontend-modern/src/utils/systemSettingsPresentation.ts must present one explicit Usage data and privacy model centered on Outbound usage telemetry; maintainer commercial-event controls, upgrade-metrics labels, and sales/onboarding reporting language must not appear in customer-facing Settings or support diagnostics, and public configuration docs must not list their internal compatibility switches as ordinary operator settings. Customer frontend code must also not import, define, or call upgradeMetrics, conversionEvents, infrastructure onboarding metrics wrappers, or POST those events to /api/upgrade-metrics/events. The telemetry copy must describe normalized release identity rather than falling back to ambiguous telemetry, upgrade metrics, or raw-version wording. Shared table, disclosure, and form primitives must also stay explicitly typed at the browser edge. Summary rows may memoize repeated pending-update reads, shared buttons must preserve discriminated disclosure props, toggle and a11y helpers must expose exact event signatures, shared rows must accept typed data-* props, and reporting-panel helpers must remain ES2020-safe instead of depending on feature-local casts or newer string helpers. That same shared settings-shell and banner boundary now also owns demo-mode commercial suppression. frontend-modern/src/components/Settings/settingsNavCatalog.ts, frontend-modern/src/components/Settings/settingsNavVisibility.ts, frontend-modern/src/stores/sessionCapabilities.ts, frontend-modern/src/stores/sessionPresentationPolicy.ts, frontend-modern/src/stores/demoMode.ts, frontend-modern/src/stores/license.ts, frontend-modern/src/stores/licenseCommercial.ts, frontend-modern/src/useAppRuntimeState.ts, frontend-modern/src/components/shared/HistoryChartOverlay.tsx, frontend-modern/src/features/patrol/PatrolIntelligenceBanners.tsx, and frontend-modern/src/features/patrol/PatrolIntelligenceHeader.tsx must consume one shared bootstrap truth from /api/security/status. The backend capability fact sessionCapabilities.demoMode remains part of that payload, but the browser-owned shared primitive is now the resolved sessionPresentationPolicy contract, which hides billing tabs, trial nudges, monitored-system warning banners, dashboard upsells, Patrol upgrade CTAs, history-lock paywalls, and other public-demo commercial affordances when the browser is rendering a public demo runtime. Platform stale-agent notices are also command-style upgrade affordances: frontend-modern/src/features/platformPage/PlatformOutdatedAgentNotice.tsx must stay hidden while the resolved presentation policy marks the session read-only, including public demo mode, even though the same notice remains available for ordinary customer installs that report outdated agents. That same shared settings-shell boundary also owns demo-mode organization suppression. frontend-modern/src/components/Settings/settingsNavigationModel.ts, frontend-modern/src/components/Settings/settingsNavCatalog.ts, frontend-modern/src/components/Settings/settingsNavVisibility.ts, frontend-modern/src/stores/sessionPresentationPolicy.ts, and frontend-modern/src/useAppRuntimeState.ts must fail closed on organization navigation and app-shell org chrome until the resolved presentation policy is known, then keep org switchers, visible Default Organization labels, and organization-scoped settings groups hidden when the browser is rendering a public demo runtime. Shared primitives must not perform their own ad hoc /api/health polling, response-header inference, hostname heuristics, or per-banner demo branching; the runtime bootstrap, shared presentation-policy store, and shared banner hooks stay on one canonical owner so suppression stays coherent across customer-facing surfaces. That same shared primitive boundary now also treats runtime capability reads and commercial reads as separate stores. Shared settings shells and banner hooks may read feature truth from frontend-modern/src/stores/license.ts, but commercial identity, upgrade routing, and trial state must stay in frontend-modern/src/stores/licenseCommercial.ts, which suppresses public-demo loads locally and defers its first fetch until the presentation policy has resolved instead of depending on route-local guards. That same shared primitive boundary now also centralizes authenticated-shell commercial posture bootstrap. frontend-modern/src/useAppRuntimeState.ts owns the first shared loadCommercialPosture() read after authenticated app runtime has mounted, while frontend-modern/src/AppLayout.tsx, frontend-modern/src/components/Settings/Settings.tsx, Patrol state hooks, and settings-panel state hooks must consume the resolved store state instead of reissuing mount-time posture fetches from each surface. Shared commercial posture loading may still dedupe or force-refresh through the store for governed billing or first-run flows, but route-local or panel-local bootstrap ownership is forbidden. Storage disk drawers now also sit on that same shared-primitives floor. frontend-modern/src/components/Storage/DiskDetail.tsx must render physical- disk read, write, and busy charts through HistoryChart plus useHistoryChartState, using the canonical physical-disk history resource id, instead of reviving diskMetricsHistory, a page-local ring buffer, or another storage-only live chart primitive for the same telemetry. That same shared-primitive floor now also owns upgrade-navigation semantics. frontend-modern/src/utils/upgradeNavigation.ts is the canonical typed internal-vs-external destination helper, while frontend-modern/src/components/shared/UpgradeLink.tsx owns how shared paywall surfaces navigate those destinations. Feature shells may request a commercial destination, but they must not re-decide whether that destination opens in-app or in a new tab once the shared primitive exists. That same shared-primitive floor now also owns prerelease shell guidance. frontend-modern/src/AppLayout.tsx is the canonical authenticated-shell owner for prerelease presentation, and the remaining user-facing treatment is the compact Preview badge keyed from resolved release metadata. Feature pages, settings panels, shared components, and route-local shells must not add a second release-candidate banner, hardcoded GitHub release or feedback links, or page-local prerelease notices once that shared shell contract exists. Browser proof for that shell rule now lives in tests/integration/tests/57-release-candidate-shell.spec.ts, which must keep rc-channel builds banner-free while preserving the compact preview badge.

The subsystem registry now also requires explicit proof-policy coverage for all shared runtime files, and shared-component guardrails fail if raw table composition is reintroduced in new shared components outside the canonical allowlist. Retained-value query ownership is now part of that shared floor too. frontend-modern/src/hooks/createNonSuspendingQuery.ts is the canonical shared helper for page-local fetches that must stay inside the mounted surface instead of falling through the app-level Loading view... fallback. Feature slices such as recovery and infrastructure drawers may consume that helper, but they must not fork new suspense-escape helpers once the shared contract exists. That retained-value boundary is explicitly bounded for long-lived browser sessions. Fulfilled resource/range queries use a 64-entry LRU with a five-minute inactive lifetime, clear on organization changes, and reject late old-scope completions from the shared cache. Large storage summary history follows the same ownership rule through frontend-modern/src/utils/storageSummaryCache.ts: it keeps at most 20 recent node/range summaries and aborts in-flight requests when organization ownership changes. Drawer navigation, chart range churn, background tabs, reconnects, and server restarts must not turn either cache into an append-only browser history. Consumers whose visible meaning changes with the source key may set retainPreviousValueOnSourceChange to false. When the new key has no retained entry, the helper must clear to the consumer's initial value in the same reactive turn before starting the replacement request; a failed or slow history range request must never leave data from the previous range labeled as the newly selected range. Cached data for the exact new key may still render immediately and refresh in the background. Each query run receives an AbortSignal; changing the source, resetting the query, or unmounting its owner must abort the superseded browser request before starting replacement work. Consumers must forward that signal through their API/cache layer when the transport supports cancellation. A poll (pollMs) does not replace a read that an earlier poll started while that read is still in flight; it skips the tick. The API answers 429 with Retry-After: 60 and apiFetch honours up to two minutes, so a polled read slower than the poll interval would otherwise be discarded on every tick and never settle. Any other read (a refresh, a source change, a remount revalidation) stays replaceable at the next tick, so a slow or hung refresh is replaced at most once before polling protects its replacement. A polled read still in flight at the first poll tick at or after STALLED_QUERY_READ_MS (150 seconds) is presumed stalled: that tick replaces it and settles it as a failed read, clearing loading, setting resolvedOnce and publishing a timed-out error, so no consumer is left waiting on a read that never returns. Each consumer renders that failure as it renders any other (the Proxmox Replication table shows its error and Retry; workload sparklines keep retained series). A transport that always takes longer than that deadline plus one poll interval never publishes. pollMs may also be an accessor, for a cadence that follows state (23 October 2026): a Patrol finding review re-reads its investigation every 5 seconds while the run is active and every 30 seconds otherwise, and its thread only while the run is live or after a failed read. A new cadence restarts the interval, a read already in flight keeps its protection, and an accessor that returns undefined stops polling. A consumer that needs two cadences uses one accessor rather than a second timer that calls refetch, because a refetch and a poll replace each other's slow reads. The settings reporting shell now also owns a deliberate split between historical performance reports and current-state VM inventory export. frontend-modern/src/components/Settings/ReportingPanel.tsx, frontend-modern/src/components/Settings/useReportingPanelState.ts, frontend-modern/src/components/Settings/reportingCatalogModel.ts, frontend-modern/src/components/Settings/reportingPanelModel.ts, and frontend-modern/src/components/Settings/reportingInventoryExportModel.ts must keep those as separate operator jobs with separate request builders and success copy, rather than collapsing inventory export back into the metrics-report controls. The schedule editor also offers vm_inventory as a current-state CSV job. It explains the whole-organization scope and new-VM discovery, hides historical format/resource/tag controls and keeps weekly/monthly cadence and email/disk delivery available. Kind switching clears stale performance selections, while saved inventory schedules retain their kind on reload and edit. Existing delivery checks, outcomes and recovery controls apply without a separate UI. That same settings shell must now also render both historical performance options and VM inventory schema from the backend-owned reporting catalog rather than hardcoding panel copy, routes, or range presets in the frontend. The frontend models may validate and present the catalog, but the canonical panel title, descriptions, endpoints, filename prefixes, range windows, and column list belong to the API reporting contract. That same settings-shell boundary now also owns operator-facing docs referrals for governed security panels. APIAccessPanel.tsx and SecurityOverviewPanel.tsx must route scope and proxy-auth guidance through the shared shipped-doc helper in frontend-modern/src/utils/docsLinks.ts instead of hardcoding GitHub main URLs that can drift from the running build, and tests/integration/tests/20-local-doc-links.spec.ts must keep browser proof on those settings-shell surfaces. That same settings-shell boundary now also owns the remediation framing for Security Overview itself. SecurityOverviewPanel.tsx may not stop at a score card and static best-practices copy once low-risk security debt has been demoted out of the global banner; it must render explicit next-step hardening actions on the canonical settings shell, source those actions from the shared security presentation owner, and keep direct operator links pointed at the owning auth, API-access, or shipped security-guide surface. The canonical proof for that shell framing remains frontend-modern/src/components/Settings/__tests__/settingsArchitecture.test.ts. The same reporting catalog ownership now also governs the operator resource- selection cap for performance reports. ReportingPanel.tsx and ResourcePicker.tsx may present or enforce that limit, but they must receive it from the backend-owned multiResourceMax definition rather than hardcoding the reporting cap in frontend-local constants. That same catalog-owned capability contract also governs which optional performance-report controls appear at all. The settings shell and reporting request builder may not assume metric filtering or custom titles are always available; they must honor supportsMetricFilter and supportsCustomTitle from the backend catalog and avoid emitting unsupported controls or request parameters from frontend-local defaults. That same backend-owned catalog also owns the initial reporting selections and transport details. useReportingPanelState.ts, reportingPanelModel.ts, and reportingInventoryExportModel.ts may not seed format/range selections from legacy frontend constants or invent fallback report endpoints, filename prefixes, default report titles, or range windows or fallback filename date-stamp styles when the catalog is present; the first valid selection and all request semantics must come from the parsed backend definition. The same rule applies to VM inventory export transport details: request builders and fallback filenames must derive the export format and extension from the parsed inventory definition instead of hardcoding csv in frontend helpers. That same fallback contract also includes the single-report filename subject, so frontend download builders may not substitute resource display names when the catalog says fallback attachment names are keyed off canonical resource IDs. That same reporting shell must also route failed catalog/report/export responses through the shared API error extractor in frontend-modern/src/utils/apiClient.ts rather than surfacing raw JSON payload text from response.text() directly in warning UI. That same reporting transport contract also means the frontend download path must prefer the backend Content-Disposition filename over any locally built fallback name when a report or inventory export response arrives. That same settings shell must also read the reporting catalog for locked users, not just entitled users, so the paywalled reporting panel does not drift onto a separate frontend-owned title or description contract. That same settings shell must now also treat the reporting catalog as the feature-identity source once it loads. ReportingPanel.tsx and useReportingPanelState.ts may use a generic loading or error shell before the catalog is available, but they must not hardcode the reporting feature key or the entitled and locked panel title and description once the catalog has loaded. The same metadata route is readable without the reporting feature gate, so the settings shell must not delay the catalog fetch on licenseLoaded() before it can render its canonical loading, locked, or entitled states. That same shell must also stay usable against older Pulse backends that do not yet expose /api/admin/reports/catalog. When that specific metadata route returns 404, useReportingPanelState.ts may fall back to the governed legacy performance-report transport (/api/reporting and /api/reporting/generate-multi) so the reporting panel does not go dead on mixed-version installs, but that compatibility path is intentionally report-only and must not invent the newer catalog-owned VM inventory export surface. reportingPanelModel.ts picks the single-resource transport from the catalog's performanceReport.singleResourceMethod: POST sends the request fields as a JSON body, which is the only transport the server lets narrate a PDF with Pulse Assistant, and GET sends the same fields as query parameters. reportingCatalogModel.ts reads a missing field as GET, the only method a backend that predates the field accepts, and the legacy compatibility catalog names GET for /api/reporting. Fleet reports stay POST on every backend. ReportingPanel.tsx must therefore treat vmInventoryExport as optional when it renders a governed reporting catalog. A legacy compatibility catalog with no inventory export still owns a valid enabled reporting surface and must continue to render the performance-report workflow instead of collapsing back to the unavailable shell. That same catalog load must also remain retryable after transient failure. useReportingPanelState.ts may memoize or dedupe in-flight work, but it must not permanently latch a failed first fetch and force operators to reload the entire settings page before the reporting shell can recover. That same contract also includes the locked teaser copy itself. The reporting catalog owns report-builder identity once the feature is available, but a locked session must render neutral feature-gate copy from the reporting panel state so Community users do not see enabled report-builder language before advanced reporting is available. That same reporting catalog also owns the enabled-shell guidance callout that explains when to use performance reports versus VM inventory export. ReportingPanel.tsx may choose the presentation primitive, but the callout title and description must come from the parsed catalog instead of a frontend-local explainer paragraph. The shared updates settings owner also defines the user-facing framing for rc-tagged builds. frontend-modern/src/components/Settings/updatesSettingsModel.ts and frontend-modern/src/utils/updatesPresentation.ts must present that channel as a prerelease or preview path with manual validation expectations, not as a near-ready release candidate promise. The root app shell now also treats backend availability as distinct from websocket liveness: frontend-modern/src/AppLayout.tsx and frontend-modern/src/useAppRuntimeState.ts must keep the top-right connection badge aligned to overall backend availability so a healthy dev/runtime backend does not present the whole shell as reconnecting just because the live stream is transiently renegotiating. That shell badge must now stay on an explicit state model as well: healthy runtime, backend-healthy-but-stream-degraded, full reconnect, and full disconnect are distinct operator states, and the shared shell may not collapse them back into one generic reconnect label. Shared feature presentation helpers under frontend-modern/src/features/ now also need to preserve route-owned page-health semantics when the owning surface is REST-backed: operators should only see reconnect or disconnected shells when the route's own data contract is unhealthy, not because a global websocket singleton is transiently reconnecting. Those same feature-owned header badges must also stay aligned to the owning runtime state instead of surfacing stale auxiliary counters as primary status. Legacy hosted-model credit counters may remain parseable on transport payloads, but shared Patrol headers must not render them as customer-facing badges in the normal self-hosted GA app. The same shared shell rule applies to retired hosted availability copy: when an older backend payload describes hosted model activation, credit exhaustion, or account-backed AI availability, shared feature shells must normalize the operator-facing guidance back to provider setup or local model setup rather than rendering the retired offer. The same primitive boundary now also owns the first AI enable control in AISettings.tsx: the primary toggle must remain explicitly addressable with a stable accessible label and route unconfigured installs into the canonical provider setup modal instead of falling back to generic "first pressed toggle" selectors, provider-model-load heuristics, legacy hosted-model enablement, or in-app trial acquisition. A settings toggle disabled by a plan or runtime lock must explain the lock in the card that owns it, not through the shared Toggle locked and lockedMessage props, which the primitive does not render. AISettings.tsx renders the Container update risk lock from the canonical Patrol availability presentation and links that explanation to the switch with ariaDescribedBy. The shared Patrol feature shells (PatrolIntelligenceHeader.tsx and PatrolIntelligenceBanners.tsx) and that card pass the session's real commercial and upgrade presentation policy into the presentation, never a constant, so a session that hides commercial surfaces gets the neutral lock line and never Pulse Pro runtime wording or a download link. That same route-owned presentation rule also governs Patrol findings empty states: shared section shells under frontend-modern/src/features/patrol/ must not render a green healthy empty state from 0 active findings alone when the owning Patrol runtime or overall-health summary is degraded, blocked, or not fully verified. The same empty-state helper must consume Patrol trust-history evidence so a historical regression reads as history review context, not as a current issue and not as a healthy all-clear. frontend-modern/scripts/canonical-platform-audit.mjs carries no Patrol summary-card rule. Its canonical-patrol/no-local-summary-card-presentation regex matched only text shaped like the deleted summary card's tint ternaries on summaryStats().criticalFindings, warningFindings and fixedCount. The Patrol state hook's summaryStats accessor has had no reader since that card was deleted, so the accessor and the rule are both gone. frontend-modern/src/utils/patrolSummaryPresentation.ts is off the audit allowlist too: it now holds only the recency and verification presenters, which no audit rule needs to exempt, so the audit scans it like any other module. The Patrol page renders no summary shell today. If one returns, the same hierarchy applies inside it: once its primary assessment strip states Patrol's current risk and verification basis, supporting metrics under that strip must stay metric-oriented and must not repeat assessment or verification labels as a second compact verdict row. Its collapsed assessment strip must remain a compact readout rather than a headline-plus-paragraph block; explanatory assessment and recommendation copy belongs in the owning Findings, Runs, or Assistant chat surfaces rather than a normal-path summary details expansion. That readout should lead with current operator state and score rather than mixing a reassuring grade label with issue-state copy in the same line. That same summary shell should also keep the shared Pulse surface neutral: severity belongs in compact accents, inline readouts, and badges rather than turning the whole assessment into a tinted warning banner, nested card, or hero-style block that breaks the surrounding operator workflow. That same summary-shell rule also applies to timing metadata: the header recency line already presents the governed Patrol activity timestamp, so a returning summary chip row must not add another recency badge that competes with it. The same default-readout rule applies to collapsed Patrol issue rows: MetadataBadge may carry severity, recurrence, and active decision/work states, but the default Patrol page must not render raw lifecycle or investigation process badges such as detected, review finding, loop state, status, outcome, or confidence as row chrome. Those details belong in expansion, run history, Assistant handoff context, or diagnostics. The same ownership split applies to supporting counts: if the Patrol summary surface renders the metric strip for active findings, warnings, criticals, and fixes, the primary assessment strip should not repeat those same counts in badge form beside the assessment and verification copy. That same ownership rule applies to empty-state timing metadata. When the Patrol page header already carries schedule and recency context, the findings empty state should not add its own footer for Last activity, Next run, or run interval text. Those supporting cards must also keep their content factual and count-based: active findings, critical findings, warnings, and fixes are valid secondary readouts, while labels such as Issues detected or Partial verification belong only to the primary Patrol assessment and verification surfaces. The same applies to Patrol operational context during active execution: the shared feature surface may add an explicit run-in-progress badge, but any activity support surface or integrated summary panel must remain factual activity copy rather than shifting into a second Patrol verdict label while a run is underway. frontend-modern/src/components/shared/TagBadges.tsx is now also the canonical tag-badge primitive. Workload rows and the unified-resource detail drawer must import that shared owner instead of keeping a workload-local tag badge variant or importing a feature-local path into infrastructure surfaces. That same owner now also holds the CSP-safe tag-dot rendering contract: tag color and active-state emphasis must travel through SVG fill/stroke attributes or stable classes, not inline background-color, box-shadow, or other style= mutations that break the hosted demo CSP. The shared tag owner also carries the complete accessibility contract for compact tag disclosure. Every rendered tag must expose its name without depending on pointer hover. Callers that provide filtering behavior must get named native toggle buttons with current selection represented by aria-pressed; informational tags and collapsed tag counts must remain keyboard focusable so the same tooltip content is available on focus. Pointer leave, blur, and Escape must dismiss those tooltips, focus must remain visibly identifiable, and the compact dot presentation must use spaced 20px targets rather than restoring an 8px click-only hit area. Feature rows and detail surfaces must extend these shared semantics instead of wrapping tag dots in feature-local click or tooltip handlers. TagBadges also owns Proxmox tag color fidelity. When a caller supplies a source instance, the primitive must read that instance's pveTagStyles entry before the legacy aggregate pveTagColors map, and it must honor the Proxmox caseSensitive flag for both override lookup and deterministic fallback color generation. Feature rows may pass instance identity into the primitive, but they must not rebuild Proxmox color-map lookup locally. frontend-modern/src/components/Settings/OperationsPanel.tsx is now also the canonical shared settings wrapper for operations-style panels such as diagnostics, reporting, and system logs. Those surfaces must extend that owner instead of rebuilding a local SettingsPanel wrapper, panel-header action slot, or divided content-body framing inline. The system logs operations surface now follows the same shell/runtime split as the other modernized settings panels: frontend-modern/src/components/Settings/SystemLogsPanel.tsx owns the operations framing and consumes the canonical stream-copy/status helpers from frontend-modern/src/utils/systemLogsPresentation.ts, while frontend-modern/src/components/Settings/useSystemLogsPanelState.ts owns the stream lifecycle, buffering, level updates, and download action. Future system logs work must extend that split instead of pulling EventSource, API calls, notification flow, or customer-facing system-log copy back into the panel render shell. Self-hosted trial CTA removal is now part of that same primitive boundary for settings and shared paywalls. Shared/settings runtime owners may derive neutral plan and entitlement posture from the canonical commercial contracts, but ordinary self-hosted feature gates must not present in-app trial starts, trial-status banners, or trial-specific rate-limit copy. Operator-facing paid handoff remains limited to explicit Plans, hosted, activation, recovery, or support surfaces; feature gates may show neutral "View plans" links through frontend-modern/src/utils/upgradePresentation.ts only where presentation policy allows commercial discovery. The shared self-hosted billing presentation and plan-section prop contract must therefore expose only the ordinary typed plan handoff. It must not retain a parallel trial label, card-required note, trial=1 destination, or dormant trial-action slot after self-hosted trial acquisition has been retired. Top-level route files are now expected to stay thin when a feature owns the real product surface, but the former /infrastructure surface is not one of those compatibility cases. It never shipped as a stable v6 route, so frontend-modern/src/App.tsx must not register it and future feature surfaces must extend Settings infrastructure, platform/runtime pages, or shared Infrastructure components instead of recreating frontend-modern/src/features/infrastructure/ as a hidden page shell. Infrastructure resource consumers may opt into websocket-first unified resource hydration only when they also schedule canonical REST revalidation after the first-paint settle window; shared route composition must not re-route the table through a blocking resource fetch just to confirm infrastructure that the realtime store has already reported. Authenticated cold starts must render from retained realtime or unified-resource state without falling back to first-run/welcome posture or replaying stale setup success notifications, and background revalidation may update rows in place but may not blank the page. Realtime resource adapters render each server row as one display row and never re-join host rows: the websocket payload, its deltas and /api/state already carry the server's presentation coalesce, which applies the operator's merge exclusions and provider scope the browser cannot see, so a browser-side join could only undo a split the server made. A row whose own server sources list shrank is rebuilt from the incoming row instead of inheriting the previous display row's provider metadata. Shared identity helpers in frontend-modern/src/utils/resourceIdentity.ts and frontend-modern/src/utils/agentResources.ts use the trimmed-string helper from frontend-modern/src/utils/stringUtils.ts so shared components do not keep their own copy of the same whitespace-trimming identity logic. Shared metric cells such as frontend-modern/src/components/shared/responsive/ResponsiveMetricCell.tsx take resolved warning/critical thresholds from the table that renders them, but threshold selection itself remains alerts-owned through frontend-modern/src/stores/alertsActivation.ts and frontend-modern/src/utils/metricThresholds.ts; shared primitive cells and rows must only pass resolved warning/critical values into metric presentation. That alerts-owned boundary distinguishes detector state from external notification delivery. Shared tables, cells, navigation, and localized copy may use detectionEnabled to decide whether in-product alert evidence is applicable, but they must not derive visibility or threshold presentation from activationState. The activation control is presented as notification delivery only, and its localized paused copy must state that detection and in-product active-alert visibility continue. Object detail drawers follow one operator-first information hierarchy across platform implementations. Overview must begin with DrawerAttentionSection when active alert or health evidence exists and show the actual problem text, not only repeat a coloured status. DrawerAttentionSection is a compact bounded alert list, not a generic detail-table card: every row preserves the backend-authored problem text, renders the actual info/warning/critical/ acknowledged severity, and reveals overflow through an in-place accessible disclosure rather than an inert hidden-count row. Aggregate drawers that include alerts from child resources must also preserve the affected resource identity and alert metric beside that text rather than collapsing them into generic VM or host labels. The Mail Gateway drawer receives its row's open alerts, matched across the gateway's identity aliases, and its header status follows them, so it never shows "Healthy" beside an active warning; its inline detail row sets whitespace-normal because platform table cells do not wrap. The row itself marks the metric an alert is about instead of stacking a reason under the name, so it keeps the shared single-line rhythm. Drawers for unified resources (Machines/Infrastructure, Docker hosts) read their open alerts from the websocket activeAlerts map through getAlertsForUnifiedResource (frontend-modern/src/utils/alerts.ts), most severe first, never from a resource-embedded list, and show each alert's resource name and metric beside its text like the other drawers. Their table rows (Docker hosts, Machines, Kubernetes nodes) take the platform host rows' red (critical) or yellow tint from the same set through getUnifiedResourceAlertStyles, only for unacknowledged alerts and never on an expanded row, so a row and its drawer agree. The Alerts > Thresholds "Active alert" dot reads the same map by each alert's resourceId and type, never by rebuilding an alert-ID key; a threshold row names the resource IDs its alerts are raised on (alertResourceIds, with getUnifiedResourceAlertResourceIds for unified guest rows) and machine rows also answer for their disks' alerts. The remaining Overview rows are additive operator context that the parent table row cannot carry, such as OS/runtime, Pulse observation or action coverage, primary reachability, protection gaps, pending updates, or an identified service. Routine health, placement, and metric values already visible in the row must not be restated merely to fill the drawer. Curated raw IDs, kernel/build strings, interface and disk facts, and provider metadata use the same compact DetailSectionTable rows as Overview; those facts must not switch to a provider-local card mosaic. Since 2026-10-14 every object drawer keeps them under its one closed Details disclosure (below). The Proxmox guest and node drawers (GuestDrawerOverview.tsx, NodeDrawerOverview.tsx) are the exception, by maintainer decision on 2026-10-13 (demand ledger "Drawer Overview leads with attention and trends", a named bet): a grid of equal-weight inventory cards buried the drawer's job. Their Overview leads with active alerts (or one "No active alerts" line), then the resource's history charts (GuestDrawerHistory, lazy-loaded) opening on 24 hours with the range picker beside them. These drawers have no History tab (nor, since the same change was carried to them, do the Docker host drawer and the shared resource-detail drawer for agents, containers and other platform rows): by maintainer decision on 2026-10-13 the tab was removed because it repeated Overview's charts at another range in another style. The charts poll only while Overview shows, and before any history is stored they show the live reading, qualified "current" or "last known". An active alert on a charted metric tones that reading and, when it carries a numeric trigger, draws it as a dashed line in its severity colour; getDrawerChartAlerts maps alert types to chart metrics and converts I/O triggers from the alert engine's MiB/s to the charts' bytes per second. Each line is a bucket average, so a tinted band behind it spans every bucket's own lowest and highest reading (buildGuestDrawerHistoryBandPath): a short spike stays visible after downsampling. While a time is inspected the readout names that bucket's peak when it rises at least a tenth of the plot above the average, and the accessible value states each reading's average and range. Every history range picker (these drawers, storage disk and pool details) names its options through formatHistoryRangeOptionLabel in historyChartModel.ts ("1 hour", "24 hours", "7 days"), so one range is never shown as "24h", "24 hours" and "Last 24 hours" in three places. Below the sm breakpoint the chart panels stack, so a phone shows the first panel (Utilization) plus any panel carrying an active alert, and one full-width "Show " button (aria-expanded, hidden from sm up) reveals the rest; wider layouts always show every panel. The charts poll every 30 seconds and offer no manual refresh; a "Retry history" button appears only while a read has failed, and when a successful retry removes it keyboard focus moves to the history region rather than the page. Whether Retry held focus is read as its branch is disposed, before the node leaves the DOM, because Chromium blurs a removed element and a blur-tracked flag would already be clear (the refresh test models that removal blur). A chart draws alert triggers only from the resource's own alerts (getOwnAlertsForUnifiedResource): a Docker host's Needs attention lists its containers' alerts, but a container's memory trigger never lands on the host's memory chart. Any warning- or danger-toned detail row is lifted into a "Worth checking" section in front, whichever section owns it, so demotion never hides a reason a reading cannot be trusted. Guest filesystems, physical disks, RAID, node GPU and NVMe sensor readings (which carry no heat tone of their own), and any availability check that is not both healthy and fresh stay visible; "availability checks passing" is said only of fresh healthy evidence. An offline or stale node says so in one line above its charts, which then take no live fallback reading, and a guest's withdrawn memory reading shows no memory size either. A node chart draws alert lines only for the node's own alerts, which ProxmoxNodesTable.tsx passes as ownAlerts by identity match; the drawer's alerts list also carries its guests' alerts through the node-name match. A one-line facts strip carries OS, the amounts the charts' percentages are of (guest vCPUs and used-of-total RAM and disk; node CPUs and used-of-total RAM and root disk, capacity only while its readings are withheld), address, a healthy last completed backup and tags. The complete inventory sits in a closed "Details" TechnicalDetailsDisclosure with keepMounted, so its rows stay in the document for find-in-page and tests while hidden. On 2026-10-14 the maintainer carried this layout to every object drawer: the shared resource-detail drawer (agents, containers, Kubernetes, TrueNAS, vSphere and other platform rows) and the Docker host drawer now open with the same alerts line, "Worth checking", charts, a facts line (container: image, Compose project and service, address; host: logical CPUs, used-of-total RAM and disk, address; Docker host: named OS, RAM and disk, the engine version being in its row) and one closed "Details" holding every inventory section, healthy availability checks and, for the resource drawer, the provider "Platform details" that had their own disclosure. Facts state used amounts only while the resource is online or running and its memory observation, if any, is current. A row is lifted when it is warning- or danger-toned or carries lift (an available container image update), unless one of the resource's own alerts already states it (getAlertStatedRowLabels, partitionDrawerExceptionRows): a container's health alert leaves "Health check: Failing" in Details rather than repeating it. An offline resource's "Observed state" is danger-toned with when it was last seen, so it is lifted without an alert. A passing availability check moves into Details, except on an endpoint that is itself a check (network-endpoint, platform availability), whose cards are its whole reading and stay in front. A container's lifted "Image update: Available" carries its next steps in place: the same shared UpdateButton as the containers table's Updates column (both built by getDockerContainerUpdateAction in features/docker/dockerContainerUpdateAction.ts, so readiness and refusal read the same) and a "View image tags" registry link. The drawer passes touchTarget, a 44px minimum height below sm; tables keep the compact pill. The control is one ContainerUpdateNextSteps element created once under the drawer's owner and reused by every rebuild of the inventory rows, so a refresh or clock tick never recreates the button or closes its review dialog. A container's row dot, State words and drawer header dot share mapDockerContainerStatus, so a degraded container with no reported state is amber in all three. That mapper also owns the restart verdict (dockerContainerRestartAttentionLabel): a running container past five restarts is warning in its dot, the sort, the phone State words, the Restarts column and the drawer's lifted Restarts row alike, never a green dot beside an amber or red count. Docker State words read in their row dot's colour through dockerStateTextToneClass (DockerNativeTableShared.tsx) with no second dot, which now delegates to the platform helper getPlatformStatusTextToneClass (features/platformPage/platformStatusTone.ts): every platform table colours the cell behind a dot's verdict through it, and PlatformIssueReason and the TrueNAS Services and VMs state notes take their tones from it too; the Docker Networks row's single name-cell dot carries its attached-container verdict the same way, and each attached-container card's state words take its dot's tone through the same helper. The Swarm tables follow it too: the cell behind a service, task or node dot's verdict (Running count and Update words, Current state, Reachability or Availability) reads in that dot's tone. "Worth checking" renders through DrawerExceptionList.tsx directly under the alerts line, above the charts: one compact label/value line per lifted row with the shared detailValueToneClass tone, not a card per section. A guest memory reading from a cache-inclusive source says "may include cache" beside its RAM amount in the facts strip; the full note (why, and what to do) is the Details row "About this reading" and appears once. DetailSectionTable keeps the single bordered table at narrow widths where density matters, then presents those same canonical rows as bounded section cards when their own container reaches 64rem. Wide-container cards share the available row width, stretch to the same row height, and use stable three- or four-column tracks. An incomplete final row must fill those same tracks with explicit integral spans; it must not independently flex-grow every remaining card into unrelated column edges. Five- and six-section drawers remain balanced across three-column rows, while seven-section drawers use a two-column span for the first card in the final row so the remaining cards keep the four-column alignment. Cards use a bounded local label column with left-aligned values so the layout has no ragged fixed-width island, stranded full-width final card, or full-drawer scan distance. Unified-resource technical summaries are part of this boundary and must not retain a full-width local table on desktop. The responsive presentation stays owned by the shared primitive; provider drawers must not fork their own desktop card renderers. Monitoring Optional detail-row progress is also owned by that shared presentation: the value text remains visible, DetailSectionTable composes ProgressBar for the bounded visual fill and accessible value, and unknown measurements render no bar. Provider drawers must not add inline width styles or local progress geometry to their section rows. mode, lifecycle, notes, maintenance, automatic-action policy, saved access configuration, and action audit belong to a dedicated Manage tab and must not render inside Overview. Guest, node, Docker-host, and unified-resource drawers must compose these shared primitives instead of defining provider-local attention or technical disclosure shells. Drawer headers reserve their limited space for the subject and the canonical full-row collapse control, plus a direct object-specific lifecycle action only where the object contract requires one. ObjectDrawerHeader owns that interaction across guest, node, Docker/unified-resource, Ceph cluster, Proxmox Mail Gateway, and shared inline detail panels: pressing anywhere in the subject row collapses the detail, Enter/Space work through the native button, and lifecycle actions do not bubble into collapse. Generic Ask Assistant and Copy context actions do not belong in object drawer headers: raw context export must not compete with the operational reading path. frontend-modern/src/components/Workloads/MetricMiniSparkline.tsx renders its hover tooltips through frontend-modern/src/components/shared/TooltipPortal.tsx rather than as HTML foreignObject shells inside the chart SVG; the mini sparkline's SVG uses preserveAspectRatio="none", where cross-browser scaling can stretch an embedded tooltip surface or drop its semantic shell styling. The retired dashboard overview route must not regain feature-local trend, KPI, problem-resource, or card shells. Workload-table and guest-row fallback copy that lives under frontend-modern/src/components/Workloads/ must keep using frontend-modern/src/utils/workloadEmptyStatePresentation.ts and frontend-modern/src/utils/workloadGuestPresentation.ts. New route-level empty states or tone mapping must extend the shared emptyStatePresentation and semanticTonePresentation helpers instead of reviving deleted dashboard-only KPI, metric, storage, recovery, problem-resource, or trend presentation helpers. That shell must also stay passive with respect to data ownership: future overview trend cards may render summary-range controls and operator-facing empty or error copy only after they have a governed owner, and they must not reintroduce route-local metrics-history fetch loops for CPU and memory sparklines; the infrastructure summary and per-workload chart routes already own that chart contract. The shared trial banner is retired for self-hosted v6 GA. Future commercial notification work must start from the explicit Plans, hosted, activation, recovery, or support surfaces rather than reviving a global authenticated-shell trial banner. The shared column picker now follows that same owner split. frontend-modern/src/components/shared/ColumnPicker.tsx stays the render shell, frontend-modern/src/components/shared/useColumnPickerState.ts owns dropdown open state and outside-click listener lifecycle, and frontend-modern/src/components/shared/columnPickerModel.ts owns hidden-column count, reset visibility policy, and column-option text-class/copy policy. Column-picker trigger badges must describe what the count means, such as N hidden, rather than exposing a bare number or ratio that competing table surfaces can interpret differently. Shared column-picker tests must cover that copy alongside the owner split so governed product tables do not regress to ambiguous utility badges. Future column-picker work should extend those owners instead of pushing document-level listener logic or column-count policy back into the shell. The shared tag input now follows that same owner split. frontend-modern/src/components/shared/TagInput.tsx stays the render shell, frontend-modern/src/components/shared/useTagInputState.ts owns input state, container-focus runtime, and tag add/remove/backspace orchestration, and frontend-modern/src/components/shared/tagInputModel.ts owns delimiter keys, placeholder policy, remove-title copy, and canonical next-tag derivation. Future tag-input work should extend those owners instead of pushing DOM reach-in or tag-mutation policy back into the shell. The shared scroll-to-top button now follows that same owner split. frontend-modern/src/components/shared/ScrollToTopButton.tsx stays the render shell, frontend-modern/src/components/shared/useScrollToTopButtonState.ts owns scroll-listener lifecycle, visible state, and smooth-scroll runtime, and frontend-modern/src/components/shared/scrollToTopButtonModel.ts owns scrollable-ancestor discovery, visibility threshold policy, aria label, and button class policy. Future scroll-to-top work should extend those owners instead of pushing scroll-container discovery or listener lifecycle back into the shell. The shared toggle now follows that same owner split. frontend-modern/src/components/shared/Toggle.tsx stays the render shell, frontend-modern/src/components/shared/useToggleState.ts owns disabled gating plus the synthetic toggle change-event runtime, and frontend-modern/src/components/shared/toggleModel.ts owns size resolution, track/knob/container class policy, and the canonical toggle event type. Future toggle work should extend those owners instead of pushing synthetic event behavior or size/class policy back into the shell. A disabled toggle (a read-only session, a plan lock, a save in flight) keeps its on or off track colour and dims, so it still shows whether the setting is on; it must not swap to a separate disabled background, which hid the track and made a locked-on setting read as off. Binary on/off controls are registry-backed too. Product surfaces must compose Toggle or TogglePrimitive for shared track/knob styling, disabled behavior, label/description wiring, and synthetic checked events instead of recreating local role="switch" buttons with aria-checked and page-local classes. Ordinary checkboxes, radio groups, and row-selection controls are separate affordances and must not be forced through this toggle primitive. The shared status badge now follows that same owner split. frontend-modern/src/components/shared/StatusBadge.tsx stays the render shell, frontend-modern/src/components/shared/useStatusBadgeState.ts owns disabled gating and click runtime, and frontend-modern/src/components/shared/statusBadgeModel.ts owns size padding, label/title fallback policy, and status-badge class selection. Future status badge work should extend those owners instead of pushing label/title policy or disabled click handling back into the shell. A badge given ariaLabel is named ": " through getStatusBadgeAriaLabel, so pages of alike On/Off badges say what each switches while keeping the visible word in the name (WCAG 2.5.3); without it the visible label remains the name. Read-only health/state badges are a separate shared primitive. frontend-modern/src/components/shared/StatusIndicatorBadge.tsx owns status-to-tone mapping, optional dot wiring, sizing, shape, and label presentation for product surfaces that display state rather than toggle it. Product components must compose StatusIndicatorBadge instead of calling getStatusIndicatorBadgeToneClasses directly; low-level status utilities may still expose the tone mapping for that primitive and utility-level tests. Platform alert severity indicators are a governed specialization of that same primitive family. frontend-modern/src/components/shared/AlertSeverityBadge.tsx owns the alert severity badge and dot shells, while frontend-modern/src/utils/alertSeverityPresentation.ts owns alert severity label formatting, severity-bucket-to-status-indicator mapping, and severity-bucket-to-detail-row tone mapping. Docker, Kubernetes, TrueNAS, vSphere, and future platform alert tables must compose AlertSeverityBadge, AlertSeverityDot, and getAlertSeverityDetailTone; they must not recreate severityVariant, severityTextClass, alertTone, or severity badge spans locally. Platform alert severity filters follow the same shared-template rule. frontend-modern/src/features/platformPage/platformAlertSeverityFilterOptions.tsx owns the canonical All/Critical/Warning/Info option labels, tones, and leading dots for platform alert table toolbars. Platform alert tables must call getPlatformAlertSeverityFilterOptions instead of declaring local severity filter arrays or calling filterChipStatusDot directly for those filters. Platform alert detail formatting follows the same by-construction primitive rule. frontend-modern/src/utils/alertDetailPresentation.ts owns the provider code labels, provider-specific resource-type labels, vSphere alert entity labels, started-at row labels, and full detail timestamp labels consumed by Docker, Kubernetes, TrueNAS, and vSphere alert tables. Those tables must call formatPlatformAlertCode, formatPlatformAlertResourceType, formatPlatformAlertEntityType, formatPlatformAlertStartedAt, and formatPlatformAlertDetailDateTime instead of declaring local formatter helpers. formatPlatformAlertStartedAt reads as an age ("2h ago"), because an open alert's Started cell answers how long it has been going on, as alert cards do; the absolute "Oct 6, 10:34 AM" also clipped at phone and tablet widths. Each table's Started cell carries formatPlatformAlertDetailDateTime as its hover title, the drawer's Started row keeps that exact time, an unset (pre-2000) or invalid time stays "-", and the column uses the numeric-value alignment kind on all four tables. Relative ages read one shared clock. useRelativeTimeNow (frontend-modern/src/utils/relativeTimeClock.ts) ticks every 30 seconds while any mounted cell reads it and stops when none do; PlatformTableRelativeTimeValue and the platform alert tables pass it as now to formatRelativeTime / formatPlatformAlertStartedAt. Platform tables keep their rows mounted across data refreshes, so an age computed only at render froze ("2h ago" three hours later) whenever its timestamp did not change. A relative-age cell must read the shared clock rather than Date.now() at render or a table-local interval. That rule covers every relative age, not only PlatformTableRelativeTimeValue cells: a value formatted through formatPlatformTableRelativeTimeValue or formatRelativeTime (compact phone ages, labelled Detail ages, row subtitles and tooltips, an open drawer's summary rows and annotations) passes now from the owning component's useRelativeTimeNow, and a countdown or freshness band derived from the same time (replication Next sync, backup age bands) reads it too. Every read of the clock returns the wall clock; the 30-second tick only tells readers to re-read, so a cell that mounts between ticks never measures from a stale time and a timestamp from the last few seconds never reads as a future time. Future times have their own formatter. formatRelativeTime is past-only: it reads a time ahead of now as "just now" (compact) or "0s ago", which is right for clock skew on something already observed and wrong for an expiry, a reminder or a schedule. Those go through formatTimeUntil (frontend-modern/src/utils/format.ts), which counts down ("in 3h", "in 1d"), rounds to the nearest unit so a duration just chosen reads as chosen, and returns dueText (default "now") once the time arrives. A countdown target does not change while its surface stays open, so the countdown passes now from the shared clock, as the Patrol suppression expiry and the Patrol findings reminder and snooze lines do. The replication Next sync column and the Patrol header's next-check CountdownTimer keep their own minute- and second-precision countdowns. The relative-age rule has one deliberate exception. The age of a latest reading (last used, last seen, last success, last checked) on data the surface reads once and does not re-read stays the age at read time, because a moving age over a snapshot that never refreshes claims the reading stopped when it may not have. Such a surface either re-reads the snapshot in the background, as Proxmox replication and the external watchdog panel do, or keeps the read-time age, as the Patrol attention detail does for Last seen. The external watchdog panel orders its re-reads by request, so a slow or hung read can neither pin it to an older snapshot nor overwrite a newer outcome while its ages keep moving. An immutable event time (when a delivery was attempted, a transition happened, a policy was set) ages correctly over any snapshot and reads the clock. Read-only metadata badges follow the same primitive-owned shell rule. frontend-modern/src/components/shared/MetadataBadge.tsx owns filled and outlined appearances, compact sizing, shape, typed tone vocabulary, fit behavior, and whitespace handling. Product surfaces such as Patrol findings may own the labels and state-to-tone mapping in their presentation helpers, but they must render visible metadata chips through MetadataBadge instead of recreating local bordered xs spans. Neutral and muted badge treatments must use semantic surface/text tokens rather than hardcoded gray palettes; non-gray typed tones may retain their state color vocabulary so success, warning, danger, info, and platform-adjacent metadata do not collapse into visually identical chips. Patrol run-history labels follow this state-badge boundary: Patrol derives the status label and typed variant in patrolRunPresentation.ts, but RunHistoryEntry.tsx must render visible state badges through StatusIndicatorBadge rather than runStatus.badgeClass or a local span. The shared segmented selector now follows that same owner split. frontend-modern/src/components/shared/FilterButtonGroup.tsx stays the render shell, frontend-modern/src/components/shared/useFilterButtonGroupState.ts owns variant resolution plus disabled selection/change runtime, and frontend-modern/src/components/shared/filterButtonGroupModel.ts owns the variant class catalog, compact-label policy, and segmented button class selection. Future filter-button-group work should extend those owners instead of pushing label truncation or segmented variant policy back into the shell. Compact labels that begin with All must preserve that scope word rather than collapsing to a trailing noun such as time; feature-owned compactLabel overrides remain authoritative when a domain needs different concise copy. Pressed/unpressed selector pills follow the same primitive rule. frontend-modern/src/components/shared/SelectablePillButton.tsx owns the pressed button shell and aria-pressed wiring, while frontend-modern/src/components/shared/selectablePillModel.ts owns the active and inactive pill class catalog. API token scope surfaces may own the security scope labels and click handlers, but must not recreate rounded-full selector pill class strings locally. A token's in-place scope editor may use semantic native checkboxes because it is a multi-select form checklist rather than a pressed pill selector; it must not imitate or fork the selectable-pill class catalog. Filter-toolbar segmented controls must delegate to this primitive rather than calling segmentedButtonClass directly, and icon+text labels must render as one inline-flex button label so compact bars keep the v5 single-line control language across Type/Status, grouped/list, bars/trends, columns, and reset. The shared selection-card primitive now follows that same owner split. frontend-modern/src/components/shared/SelectionCardGroup.tsx stays the render shell, frontend-modern/src/components/shared/useSelectionCardGroupState.ts owns variant resolution plus disabled selection/change runtime, and frontend-modern/src/components/shared/selectionCardGroupModel.ts owns the tone fallback, group/button class catalog, and title/description presentation policy. Future selection-card-group work should extend those owners instead of pushing tone or active-card presentation logic back into the shell. The shared dialog now follows that same owner split. frontend-modern/src/components/shared/Dialog.tsx stays the render shell, frontend-modern/src/components/shared/useDialogState.ts owns focus trap, body-scroll lock, previous-focus restoration, shared blocking-dialog visibility, and backdrop-close runtime, and frontend-modern/src/components/shared/dialogModel.ts owns focusable-element lookup plus layout and panel class policy. Future dialog work should extend those owners instead of pushing focus-trap lifecycle or layout policy back into the shared shell. App-shell consumers such as frontend-modern/src/App.tsx and frontend-modern/src/AppLayout.tsx may read that shared blocking-dialog state to suppress background affordances, but they must not reimplement their own parallel modal-stack bookkeeping. The shared history chart now follows the same owner shape. frontend-modern/src/components/shared/HistoryChart.tsx stays the render shell, frontend-modern/src/components/shared/useHistoryChartState.ts owns license gating, history fetch/refresh, canvas draw lifecycle, and hover state, and frontend-modern/src/components/shared/historyChartModel.ts owns tooltip formatting, scale and axis math, and closest-point selection. Lock overlays in ordinary self-hosted surfaces must stay informational rather than presenting trial-start or upgrade-link actions. Future history-chart work should extend those owners instead of pushing fetch, license, commercial trial actions, or canvas math back into the shared component shell. The shared history range catalog is also owned here. The canonical product range sequence is 24h, 7d, 14d, 30d, and 90d, with 14d preserved as the Relay entitlement surface rather than hidden behind the Pro-only long-range controls. Lock copy must derive its target days and tier label from the selected range instead of assuming every locked history selection is a 30-day or 90-day Pro ask. The remaining header, overlay, and tooltip render surfaces now live in frontend-modern/src/components/shared/HistoryChartHeader.tsx, frontend-modern/src/components/shared/HistoryChartOverlay.tsx, and frontend-modern/src/components/shared/HistoryChartTooltip.tsx instead of re-accumulating those sections inline in the shell. That tooltip owner now also holds the CSP-safe hover contract: chart tooltips must render inside the chart surface with model-owned layout and SVG/attribute positioning, not through fixed portals or inline left/top style attributes that violate the public demo CSP. The same CSP rule binds every runtime component, not only these owners. Solid compiles a static style value (a style="..." string, or any literal entry of a style={{ ... }} object) into the element's template HTML, and the production style-src 'self' 'nonce-...' policy reports each one as a style-src-attr violation when that template is parsed. Static styling must be a class: a Tailwind utility, or an arbitrary property such as the [overflow-anchor:none] on drawer tab panels. Only dynamic values may use style, because Solid applies those through CSSOM. frontend-modern/src/components/shared/SharedPrimitives.guardrails.test.ts compiles every runtime .tsx that mentions style with the client-build Solid preset and fails on any style attribute or <style> element in the resulting templates, including nested <template> content. Tooltip shell chrome must follow semantic surface, text, and border tokens rather than hardcoded dark palette utilities so light and dark themes share one primitive-owned contrast contract. The shared container update badge now follows that same owner split. frontend-modern/src/components/shared/ContainerUpdateBadge.tsx stays the render surface for the badge, icon, and update button shells, frontend-modern/src/components/shared/useContainerUpdateButtonState.ts owns Docker update mutation flow, persistent update-store state, settings gating, and button lifecycle, and frontend-modern/src/components/shared/containerUpdateBadgeModel.ts owns badge and button tooltip formatting, class selection, and label/state presentation. Future container-update work should extend those owners instead of pushing store wiring, settings reads, or mutation flow back into the shared shell. For governed container updates, the row's current-session pending state is bound to the action ID. A changed registry update badge, a legacy command for the same container, or elapsed time cannot turn an unconfirmed action green or offer a fresh plan. The row reopens that action with a read-only action GET; only its recorded completed outcome earns the transient completed state. Operator-closed unknown outcomes remain reviewable and must not be presented as failed. A browser reload loses this row-local shortcut, not the durable Actions audit; the audit remains the authority before any later retry. The shared web interface URL field now follows that same owner split. frontend-modern/src/components/shared/WebInterfaceUrlField.tsx stays the render shell, frontend-modern/src/components/shared/useWebInterfaceUrlFieldState.ts owns metadata fetch/save/remove lifecycle and success/error state. frontend-modern/src/components/shared/webInterfaceUrlFieldModel.ts owns URL validation and target-label normalization. The shared primitive now also supports an embedded mode with a caller- owned title so feature drawers can place web-interface controls inside a larger access surface without forking the save/remove/runtime behavior. Future web-interface URL work should extend those owners instead of pushing metadata transport or validation back into the shared shell. WebInterfaceUrlField.tsx renders read-only when sessionCanWriteMonitoringMetadata() is false. That mirrors the backend metadata write gate: browser sessions always pass, while API-token sessions such as kiosk links need monitoring:write or the wildcard scope. Read-only sessions keep the saved URL with its open and copy controls and a one-line explanation, but get no input, Save, or Remove controls that the server would refuse. Saved web-interface launch affordances must also stay on a shared primitive instead of page-local table columns or one-off external-link anchors. frontend-modern/src/components/shared/WebInterfaceLink.tsx owns URL classification, the distinct adjacent launch control, new-tab safety attributes, row-click/key propagation containment, invalid-URL presentation, and accessible launch labels. The resource name remains inert identity text; the launch control is a separate keyboard-focusable target with at least a 24-by-24 CSS-pixel hit area so activating it never doubles as row selection or drawer navigation. Workload guest rows, grouped node headers, standalone machine rows, Proxmox node rows, Docker/Podman rows, Kubernetes rows, unified host/PBS/PMG rows, and alert resource rows/group headers compose that primitive for every comparable overview surface. Missing URLs render no fake control. Malformed or non-HTTP(S) saved values render a non-interactive accessible warning rather than disappearing or becoming executable. Runtime/platform tables must not add separate Web columns, page-local external-link anchors, linked resource names, or duplicated new-tab safety handling for that launch affordance. Docker host rows follow the same primitive-owned contract end to end. DockerHostsTable.tsx composes ResourceNameWithWebInterfaceLink beside the inert host name, while the host drawer's Manage access surface embeds WebInterfaceUrlField with metadataKind="docker-host" and the stable host source id. The drawer must return saved URL changes to its owning table row so the adjacent launch control updates immediately; neither the table nor drawer may fork metadata transport, URL validation, or new-tab behavior into a Docker-local implementation. Shared-template drift enforcement is registry-backed: frontend-modern/scripts/shared-template-registry.json is the canonical list of standardized repeated affordances, required consumers, and forbidden local patterns, while frontend-modern/scripts/shared-template-audit.mjs enforces that registry. Future repeated-affordance migrations must add or extend a registry rule as part of the same change that extracts or adopts the shared primitive. Bans on local patterns live in that registry or in frontend-modern/eslint.config.js, not in tests that read component source with ?raw or readFileSync and assert a string is absent. frontend-modern/scripts/source-text-test-audit.mjs holds each test file's source-read count to its baseline, and a new test file starts at zero. Button-styled commercial upgrade CTAs are one of those registry-backed templates. frontend-modern/src/components/shared/UpgradeLink.tsx owns UpgradeButtonLink, and the registry requires gated settings, audit, agent profiles, and self-hosted plan CTA surfaces to compose it rather than styling UpgradeLink or anchors locally. Commercial presentation helpers may own the label and destination intent; they must not own CTA button chrome. Platform table frames are one of those registry-backed templates. frontend-modern/src/features/platformPage/sharedPlatformPage.tsx owns PlatformTableShell, including the canonical table card, header row, and body divide styling, plus the single-line 32-pixel summary-row rhythm inherited by Proxmox, Docker / Podman, Kubernetes, TrueNAS, vSphere, Standalone, Workloads, and direct Storage consumers. Inline detail rows remain content-sized. Secondary context must move to an existing column, a supplemental tooltip, or the inline detail drawer instead of stacking a second visual line. The shell also owns the final responsive table-class composition, so feature tables may supply breakpoint floors but cannot accidentally override the shared phone-width floor with min-w-full. Platform table frames now have no local-frame exceptions in shared-template-registry.json: new and existing platform tables must compose PlatformTableShell instead of recreating the TableCard, header row, or body divide frame locally. Platform table consumers must preserve the owner split: frontend-primitives owns the repeated PlatformTableShell frame and guardrail registry, while platform and unified-resource consumers own the source-specific row fields, drawers, and resource semantics. The rendered phone contract now applies consistently to Proxmox, Docker, Kubernetes, TrueNAS, vSphere, Standalone, and direct Storage data tables. At less than 360 pixels of content width, the shared platform-table container rule gives identity 40 percent and hides only consumer-marked platform-table-narrow-hidden cells; ordinary phone widths retain the richer five-to-seven-column projection. Summary rows remain single-line, and any value that must truncate is backed by a touch- and keyboard-operable inline detail row that exposes the complete value. Standalone Pulse Agent and Availability consumers may compose one compact status summary directly above that shared table frame. The consumer owns the already-loaded resource counts, freshness-aware attention ordering, and settings action; the shared primitive boundary still forbids a second fetch, detached proof strip, decorative chart, or locally recreated table frame. The shared platform toolbar owns count grammar and canonical status-route normalization: one visible row uses the singular form, and legacy provider status values such as running or stopped normalize into the page's shared health filter rather than leaking provider vocabulary between platform routes. Platform table empty states follow the same registry-backed ownership. PlatformTableEmptyState owns the repeated table-card empty-state shell for Docker, Kubernetes, Proxmox, Standalone, TrueNAS, vSphere, and future platform feature tables; source-specific consumers own only the empty-state icon, title, description, and actions. Platform feature tables must not import EmptyState directly or recreate a Card-wrapped empty-state shell. Embedded settings and patrol panel empty states follow the same shared-template registry, but compose EmptyState directly with variant="panel" when they are not platform table/card empty states. The primitive owns compact spacing, icon treatment, text hierarchy, framed-versus-panel density, and action-slot layout; feature panels own only the empty-state copy, icon choice, and callbacks. Migrated panels such as Agent profiles, Audit Webhooks, Audit Log, Availability, Diagnostics, SSO providers, and Patrol Run History must not recreate local text-center icon stacks, dashed empty-state frames, or page-local empty-state action buttons. Platform table loading states are registry-backed too. PlatformTableLoadingState owns the repeated table-card compact role="status" loading row for platform pages and tables; platform consumers own only the title and description copy. Platform feature surfaces must not recreate the TableCard plus compact status-row shell locally. Platform table text-cell fallback formatting is a shared primitive as well. formatPlatformTableTextValue in frontend-modern/src/features/platformPage/sharedPlatformPage.tsx owns the trimmed-string plus canonical empty-cell marker behavior. Kubernetes platform tables must compose that helper instead of declaring local textValue helpers or inlining asTrimmedString(...) || '—' fallback expressions for table text cells. Platform table title-case fallback formatting follows the same rule. formatPlatformTableTitleCaseValue owns the repeated trimmed-string plus Unknown fallback behavior for state/status labels that need simple title case. TrueNAS platform tables must compose that helper instead of declaring local titleCase helpers. The apps, shares, services and VMs tables are its required consumers. The storage table left that list when its Health cell switched from a title-cased risk-level pill to TrueNAS's own reason text, and the scope-wide guard against a local titleCase helper still covers it. Platform table compact list summaries follow the same rule. summarizePlatformTableValues owns the repeated trimming, empty-marker label, visible-value count, +N overflow suffix, full-title text, and normalized value-list behavior for dense platform table cells. Kubernetes service/network/config/policy/autoscaling tables, vSphere datastore/network tables, and TrueNAS network-share tables must compose that helper instead of declaring local compactList or summarizeValues helpers. Platform table uptime formatting follows that rule too. formatPlatformTableUptimeValue owns the repeated compact/full uptime label selection plus canonical empty-cell marker behavior for dense platform table cells. Docker / Podman hosts, Kubernetes nodes, Proxmox nodes and backup server rows, Proxmox Mail Gateway instance and drawer node rows, Standalone machines, TrueNAS systems, and vSphere ESXi hosts must compose that helper instead of declaring local formatUptime helpers or importing the generic formatter in table files for the same days/hours/minutes fallback. Platform table byte-size formatting follows the same rule. formatPlatformTableBytesValue owns the repeated positive-byte formatting plus canonical empty-cell marker behavior for dense platform table cells. Docker / Podman native storage cells, Docker / Podman engine storage-usage cells, Kubernetes node and storage capacity cells, Proxmox backup server, Ceph, coverage, and recoverable-artifact size cells, and TrueNAS system, VM, storage-topology, and protection byte cells must compose that helper instead of declaring local formatBytes wrappers, importing the generic formatter in table files, or reimplementing byte-unit precision there. Compact platform table timestamps follow the same rule. PlatformTableDateTimeValue and formatPlatformTableDateTimeValue own compact date-time parsing, invalid/empty markers, optional minimum-year filtering, Intl format options, and tabular-number styling for dense timestamp cells. TrueNAS protection completed-time cells must compose that primitive instead of declaring local compact toLocaleString helpers, and timestamp columns use the canonical numeric-value alignment kind because they are scannable scalar values. vSphere activity "When" cells read as an age instead: the absolute "Oct 4, 11:22 PM" clipped in its column at most widths, and recent vCenter activity is scanned by how long ago it happened, as Kubernetes events are. They compose PlatformTableRelativeTimeValue with the exact time in the cell's hover title and in the row drawer, and a vCenter time before 2000 (its unset marker) stays an empty marker rather than a decades-old age. Relative timestamp-age cells follow the same rule. PlatformTableRelativeTimeValue and formatPlatformTableRelativeTimeValue own the repeated formatRelativeTime composition, compact-label default, invalid/empty markers, and tabular-number styling for dense platform table cells. Docker / Podman volume created-at cells, Kubernetes deployment age cells, Kubernetes event observed-time cells, Proxmox backup created-age cells, Proxmox replication last-sync cells, Standalone machine last-seen cells, Standalone availability-check checked-at cells, and vSphere activity "When" cells must compose that primitive instead of importing formatRelativeTime directly or declaring local timestamp-age helpers in table files. Duration and interval cells follow the same rule. PlatformTableDurationValue and formatPlatformTableDurationValue own seconds/minutes/hours duration labels, explicit fallback text, canonical empty-cell markers, and tabular-number styling for dense platform table cells. Proxmox replication last-duration cells and Standalone availability-check poll interval cells must compose that primitive instead of declaring local seconds/minutes helpers. Responsive platform table width normalization is registry-backed too. getPlatformTableWeightedColumnWidthStyle owns visible-column weight normalization, zero-width fallback, and stable four-decimal percentage formatting for dense platform table models. Docker / Podman container columns and Proxmox host columns must keep their domain column IDs, layout breakpoints, and weight maps local, but must call that shared helper instead of declaring local formatPercentage / toFixed(4) width helpers. Platform table numeric fallback rendering is registry-backed too. PlatformTableNumberValue owns finite-number checking, tabular-number styling, custom empty-marker support, and caller-owned number formatting for dense optional numeric table cells. Docker / Podman native count helpers, Kubernetes optional count cells, Docker Swarm service desired/running counts, Kubernetes Deployment replica counts, Proxmox Mail Gateway count columns, and TrueNAS system share/service and storage-topology disk count cells must compose that primitive instead of declaring local numberValue, numericValue, replicaCount, countCell, diskCountLabel, or cell-level tabular-nums variants. If a scheduler, service-domain, or inventory count is intentionally zero-defaulted, the consuming table owns that field/default choice and still renders through PlatformTableNumberValue. Locale-formatted integer count labels share the same primitive boundary. formatPlatformTableIntegerValue owns rounded integer formatting, locale grouping, and empty-marker behavior for dense platform table and drawer count cells. Kubernetes namespace drawers, Proxmox Backup Server backup counts, Ceph pool object counts, and Proxmox Mail Gateway table/drawer counts must compose that helper, usually through PlatformTableNumberValue, instead of declaring local formatInteger, formatLocaleCount, formatNumber, or direct toLocaleString() count formatting. PlatformTableCountRatioValue owns the companion healthy/total or ready/total count-ratio skeleton: numerator, slash, muted denominator, tabular styling, and empty marker behavior. formatPlatformTableCountRatioValue owns the same zero-default and suffix behavior for string-only table summaries and titles. Kubernetes cluster child counts compose the component instead of keeping a table-local childCountCell renderer, and Kubernetes networking endpoint summaries compose the formatter instead of hand-building 3/3 ready strings; the consuming table owns only which current/total values, suffix, and warning tone apply. One-decimal percent and positive Celsius cells are also shared platform-table value primitives. PlatformTablePercentValue owns percent formatting, tabular-number styling, and empty markers; formatPlatformTablePercentValue owns the same one-decimal percent string for overlay labels, titles, and sparkline labels, including caller-selected ratio normalization and clamping. PlatformTableTemperatureValue owns finite positive Celsius validation, one-decimal °C formatting, tabular-number styling, and empty markers. Docker / Podman host, Proxmox backup/Ceph/node/mail-gateway, and TrueNAS system/storage-topology tables or drawers must compose those primitives instead of carrying local formatPercent, formatPercentLabel, toFixed(1)%, or temperature label helpers. Platform table metric fallback rendering is also shared. PlatformTableMetricFallback owns the centered muted empty marker used in metric bar cells plus optional caller-owned fallback label/title text, and getPlatformTableFiniteMetric owns finite-number normalization for CPU and memory, disk, and capacity values. Docker / Podman, Kubernetes, Proxmox, Standalone, TrueNAS, and vSphere platform tables and their table-model helpers must compose those helpers instead of declaring local metricFallback / finiteMetric helpers or inlining centered muted dash fallback markup in metric cells. A row whose status indicator is danger (offline) blanks every reading from its last report, not only the metric bars. Proxmox node rows already gate uptime and temperature on online. The Machines table gates its Uptime and Temperature cells, and the Docker hosts table its temperature cell, on the same check. A machine another source keeps up stays rendered, since its cell may carry that source's current reading. Numeric cells keep their own right-aligned empty dash; the centred PlatformTableMetricFallback marker stays specific to metric-bar cells. Platform table metric severity coloring is alert-backed, not hardcoded. The Docker host and container, Proxmox node, Kubernetes cluster and node, TrueNAS system and app, and vSphere host tables must resolve display thresholds through the alerts subsystem's activation store (getMetricThresholds with the platform's runtime scope and override identity candidates) and pass them into the shared metric bar primitives (ResponsiveMetricCell, StackedMemoryBar, StackedDiskBar); the static METRIC_THRESHOLDS display constants remain fallback-only presentation for callers without alert configuration in scope. Workload guest bars pass the guest's alert policy tags (getWorkloadAlertPolicyTags) through that same getMetricThresholds call, so a Proxmox guest tagged pulse-relaxed colours CPU, memory and disk by the relaxed floor the engine applies to its configured thresholds; bar primitives never read tags themselves. The Alerts thresholds table's shared row and mobile card render the relaxed-trigger line from getAlertResourceRelaxedTriggerSummary, not from per-section copy. The vSphere ESXi host table keeps power state distinct from aggregate resource health without spending a phone column on the repeated normal case. At phone widths the Power column is hidden and a shared MetadataBadge appears beside host identity only for Off, Suspended, or Unknown; powered-on hosts retain only the canonical health dot there. Tablet and desktop widths keep the full Power column, and the mobile exception badge must not replace or recolor the health indicator. Canonical Linux memory usage-unavailable is a first-class metric fallback, not a numeric zero. Shared workload bars, platform tables, drawers, and live history labels must render N/A (while retaining known capacity where useful) and must not synthesize a zero-width used segment or a healthy-looking 0%. When unified-resource metadata carries an unavailable raw Proxmox, agent, or Docker memory facet alongside a trusted metric from another source, the trusted merged metric wins; the raw facet remains diagnostic evidence only. Platform load-failure states are registry-backed as well. PlatformErrorState owns the repeated table-card error shell, warning icon, and Refresh action for platform page and table load failures; platform consumers own only the failure title, description, and refresh callback. Platform feature surfaces must not recreate an EmptyState plus local Refresh button for Could not load... states. Platform section tabs are registry-backed too. PlatformSectionTabs owns the workflow tab shell, hidden-single-tab behavior, active-link styling, link targeting, active-page aria state, and minimal active-tab visibility scrolling after route or viewport-size changes; platform page surfaces own only tab specs, the active tab choice, and aria-label copy. The shared tabs also stay unmounted in kiosk mode, so a page surface must not add its own kiosk check around them. The visibility behavior is a shared horizontal-rail boundary: horizontalRailVisibilityModel.ts owns the bounded minimal-scroll calculation, and useActiveHorizontalRailItemVisibility.ts owns route-state, resize, and rail-resize synchronization. Platform section tabs, Alerts mobile navigation, and future horizontally scrolling destination rails must compose that owner so the selected destination cannot remain clipped after direct navigation or a viewport change. Platform feature surfaces must not rebuild local nav tab bars with aria-current and border-tab styling. Filter bars are registry-backed too. FilterBar owns resource-list filtering as a catalog of FilterDef entries, while filterChipStatusDot owns the small leading status-dot glyph used by filter options. Page and feature surfaces must not copy the chip-dot <span> factory or import the legacy PageControls deck for resource-list filtering; those drift checks live in shared-template-registry.json and run through shared-template-audit.mjs. Status indicator dots are registry-backed too. StatusDot owns the shared size, color-token, pulse, title, aria, and decorative-status behavior for resource and health dots, while feature owners supply only the status semantics. Storage linked-disk health rows must derive a StatusDot variant through getLinkedDiskHealthDotVariant and must not recreate local rounded green/yellow span classes in storage components or storage-backup presentation helpers. Loading indicators are registry-backed too. LoadingSpinner owns the shared border-based spinner shell, size catalog, tone catalog, decorative status, and accessible status label behavior. Shared primitive internals such as Button, PulseDataGrid, and HistoryChartOverlay, plus Login, Settings, Patrol, and AI finding surfaces, must compose that primitive for pure loading and action-pending spinners; icon-specific refresh rotation remains local icon state, not a loading-spinner shell. Native select controls are registry-backed too. FormSelect owns label/id wiring, helper-text description merging, value synchronization, default select chrome, dynamic-option value synchronization, and compact styling hooks for native selects. Product components and shared filter/menu internals must compose FormSelect rather than recreating screen-reader labels, native <select> shells, value-reapply effects, or compact select chrome locally; the only raw native select in frontend runtime code should live inside that primitive. Native textarea controls follow the same contract. FormTextarea owns label/id wiring, helper-text description merging, value synchronization, default textarea chrome, and compact styling hooks for multi-line text fields. Alert destination fields, incident notes, infrastructure merge reports, commercial recovery input, and agent-profile prompt/description fields must compose FormTextarea; alert, settings, and infrastructure runtime code must not recreate raw native <textarea> shells outside that primitive. Search controls are registry-backed too. SearchField owns simple search input chrome, clear affordance, keyboard forwarding, focus handling, aria labels, and trailing-control padding, while SearchInput owns resource-list search enhancements such as history, tips, and type-to-search wiring. Product surfaces must compose those primitives instead of rendering native type="search" inputs or recreating search icon/clear/input classes locally. Settings resource selectors must use SearchField for both primary text search and secondary tag/resource filters instead of restoring native type="text" filter inputs beside a shared search field. Segmented selectors are registry-backed too. FilterButtonGroup owns the settings, prominent, compact, and equal segmented selector shells, including active-button tone, disabled-option behavior, pressed-state semantics, compact labels, and horizontal scroll treatment through the shared shell/state/model split. Settings and compact feature surfaces must compose that primitive instead of copying active-button selector styling locally. Selectable pill buttons are registry-backed too. SelectablePillButton owns rounded pressed/unpressed selector pills, including active tone, disabled treatment, focus ring, and aria-pressed; settings and security surfaces must compose that primitive instead of copying rounded-full active selector styling. ResourcePicker report-domain filters are part of that boundary: the picker owns the reportable resource categories and labels, but the type selector shell must come from FilterButtonGroup. Column visibility controls are registry-backed too. ColumnPicker owns the column chooser trigger, panel title, reset action, empty-state copy, hidden count badge, dropdown width, and outside-click lifecycle through the shared shell/state/model split. Table surfaces must compose that primitive instead of recreating local column chooser buttons or panels. Selection-card groups are registry-backed too. SelectionCardGroup owns the compact/detail card grid, active-card tone, disabled selection behavior, pressed-state semantics, title/description styling, and icon container treatment through the shared shell/state/model split. Settings and feature surfaces must compose that primitive instead of copying compact/detail border-card styling locally. Grouped/list table-mode controls are registry-backed as well. GroupedTableModeSegmentedControl owns the shared Group by label, Grouped / List labels, tooltip titles, and icons for table mode switching. Resource surfaces must compose that primitive instead of copying grouped/list segmented-control labels locally. Product table cards are registry-backed too. TableCard owns the shared bordered, no-padding, card-tone table frame, while TableCardHeader owns the title/action/clear chrome, clear button copy, aria label, and propagation containment for table-card headers. Product table surfaces must compose those primitives instead of recreating local overflow-hidden bordered wrappers or retired summary-table header aliases. Inline detail table rows are also registry-backed. InlineDetailTableRow owns the row/cell/content shell and row-click containment for platform, workload, and infrastructure inline drawers; callers may pass row-specific data-* attributes and colspan, but not a cell or content class, and they must not recreate the surface-alt detail row shell locally. Five tables (Proxmox Ceph, Mail Gateway, Replication and Coverage, Docker Networks) once replaced the content class to change padding and silently lost the phone width cap below, so the shell is no longer overridable; layout a drawer needs belongs on an element inside it. The content shell must clip horizontal paint below the large breakpoint without becoming a scroll container, reset the parent table's whitespace-nowrap inheritance, and allow its descendants to shrink, then restore visible overflow for the static desktop layout. Below that breakpoint the default content shell is capped at the table scroll shell's content-box inline size (max-w-[100cqi]; the shell is a size container), never a viewport estimate of the page chrome, so a phone drawer does not leave an empty strip beside its content (the retired 100vw-3.5rem cap left 27px at 390px) and is never wider than the table's visible area. Fixed-layout cells clip overflow, so the sticky offset is inert and horizontal scrolling still moves the drawer with its cell. Long operator-state copy must wrap inside the shared row border instead of painting beneath adjacent controls or disappearing at the clip edge. When focused detail content is removed, InlineDetailTableRow restores focus to its current aria-controls disclosure with preventScroll; live refresh, collapse, and row replacement must not move the surrounding application viewport merely to reveal that control. The detail cell spans the summary row's visible cells, and InlineDetailTableRow re-measures that span while the row stays open: whenever summary cells are added, removed, or hidden (a column picker toggle, or a data-driven column such as the vCenter column that shows only with several vCenters), and whenever the caller's requested colspan changes. An open drawer must never keep a phantom column or stop short of a newly shown one. Changes inside a cell's content do not re-measure, so live metric updates stay cheap. When the span grows, the row changes the cell's own width for one forced layout and restores it, because Chromium otherwise keeps a fixed-layout table's old column widths for the widened cell. Callers keep passing their requested colspan; they must not re-key or remount the detail row to force a new span, which would discard drawer state. Inline detail section content is registry-backed separately from the row shell. DetailSectionTable, InlineDetailPanel, and detailSectionModel.ts own detail row compaction, section-table rendering, value-tone classes, and the inline collapse action for platform alert/activity/protection/service detail panels. InlineDetailPanel composes ObjectDrawerHeader, so its entire heading row closes the panel instead of presenting a separate text button. It heads a record drawer the way object drawers are headed: DrawerSubjectHeading with the record's subject (the affected resource, service or dataset, never the panel kind such as "Alert detail") and its status dot, a one-line summary, and an optional actions slot for a link to the owning object; the collapse label names that subject rather than an id. Record values wrap (DetailSectionTable wrapValues), because a record is mostly sentences and identifiers that a phone cannot hover to read. DrawerSubjectHeading is the one subject heading for every drawer: status dot then name, the name wrapping rather than truncating, and the dot omitted only for a record without a status. ResourceDetailDrawer composes it too, with its source and identity badges in the trailing slot, and takes the table row's own statusIndicator through PlatformResourceDetailTableRow so the drawer dot cannot disagree with the row dot it opened from; consumers may own the platform-specific section data, but they must not recreate local DetailField grids or route platform-neutral detail tables through a provider-named primitive. Long identifying labels can opt into DetailRow.layout: 'stacked': the shared renderer places the complete label above its value and progress bar in one full-width cell on both narrow and desktop layouts. The label wraps even an unbroken path segment rather than inheriting table nowrap/ellipsis; it must be readable without hover or horizontal scrolling. Compact two-cell rows remain the default. This is a presentation choice only: value formatting, unknown usage, tone and accessible progress metadata retain their existing semantics. Platform row-detail disclosure controls are also registry-backed templates. frontend-modern/src/features/platformPage/PlatformResourceDetailTableRow.tsx owns PlatformResourceDetailToggleButton, which composes SummaryRowActionButton for the canonical row-detail affordance, accessible label, aria-expanded, aria-controls, and propagation containment. Platform tables that use createPlatformResourceDetailState or render local inline detail rows must compose that toggle; they may still own the detail row content, drawer payload, post-success refresh callbacks, and platform-specific fields. The shared help icon now follows that same owner split. frontend-modern/src/components/shared/HelpIcon.tsx stays the render shell, frontend-modern/src/components/shared/useHelpIconState.ts owns open state, popover-position lifecycle, and global click/escape listeners, and frontend-modern/src/components/shared/helpIconModel.ts owns help-content resolution, icon sizing, missing-content warnings, and popover-position math. Future help-icon work should extend those owners instead of pushing registry lookups or DOM listener lifecycle back into the shared shell. The shared mobile nav now follows that same owner split. frontend-modern/src/components/shared/MobileNavBar.tsx stays the render shell, frontend-modern/src/components/shared/useMobileNavBarState.ts owns the mutually exclusive platform/overflow menu state, keyboard focus return, outside-click dismissal, last-active-platform continuity, and click handoff lifecycle, and frontend-modern/src/components/shared/mobileNavBarModel.ts owns platform and utility tab ordering, the dedicated platform-switcher projection, alert badge counts, and tab button class policy. Future mobile-nav work should extend those owners instead of pushing tab-order or DOM lifecycle logic back into the shared shell. With support/admin controls moved under Settings, that utility ordering must no longer reserve a standalone operations slot; alerts, Patrol, and Settings are the remaining authenticated utility tabs. Platform switching is a first-level mobile action. The first bottom-rail slot must show the current or most recently active platform and open the dedicated platform menu containing every evidence-admitted platform; platform choices must not be buried in the generic More menu. Alerts, Patrol, and Actions remain pinned daily-operation destinations. More is reserved for secondary utilities such as Settings, keeping the bottom rail to five predictable targets without making infrastructure switching a two-level navigation task. The shared command palette now follows that same owner split. frontend-modern/src/components/shared/CommandPaletteModal.tsx stays the render shell, frontend-modern/src/components/shared/useCommandPaletteState.ts owns query state, selected-row keyboard state, open-reset/focus lifecycle, route-path wiring, and command selection, and frontend-modern/src/components/shared/commandPaletteModel.ts owns canonical command construction plus query normalization and filtering policy. Future command-palette work should extend those owners instead of pushing route construction or search policy back into the shared shell. The palette search composes SearchField as an editable list-autocomplete combobox: DOM focus remains on the search input while arrow-key selection is exposed through aria-activedescendant, and the active listbox option remains scrolled into view. Palette options stay outside the Tab sequence; filtering to an empty result collapses the combobox and clears its active descendant. The OpenCode reference for this interaction is packages/opencode/src/cli/cmd/run/footer.command.tsx at origin/dev e82542b8023a8374f29c23b70ec019c8f256354e, where RunCommandMenuBody builds and filters command rows, holds selected menu state via createFooterMenuState, resets selection as the query changes, and routes keyboard movement and selection through handleKey. Pulse adapts that contract by keeping command construction in the shared palette model and command execution in the shared palette state instead of moving editor-specific command routing into the render shell. The shared search field now follows that same owner split. frontend-modern/src/components/shared/SearchField.tsx stays the render shell, frontend-modern/src/components/shared/useSearchFieldState.ts owns focused- Escape clear/blur behavior and input-ref lifecycle, and frontend-modern/src/components/shared/searchFieldModel.ts owns clear/shortcut visibility rules plus trailing-control padding policy. Future search-field work should extend those owners instead of pushing event behavior or layout policy back into the shared shell. Forwarded keyboard and blur events must preserve native browser event getters and methods while normalizing currentTarget and target; shared search-field wrappers must not proxy native event properties or methods through a receiver that can break KeyboardEvent/FocusEvent getters, preventDefault(), or stopPropagation() in live browser surfaces. The shared search input now follows that same owner split. frontend-modern/src/components/shared/SearchInput.tsx stays the render shell, frontend-modern/src/components/shared/useSearchInputState.ts owns input-ref lifecycle, type-to-search registration, and enhancement runtime composition, and frontend-modern/src/components/shared/searchInputModel.ts owns the shared search-input contract plus shortcut-hint and trailing-control policy. Future search-input work should extend those owners instead of pushing type-to-search or enhancement wiring back into the shared shell. Infrastructure-aware completion follows that owner split. Product surfaces provide safe canonical identity projections through SearchInput suggestions; useSearchInputEnhancements.ts ranks those identities and exposes only the single dimmed inline suffix after the current query. It must not introduce a second results dropdown. When several identities match, the suffix stops at their unambiguous common prefix instead of selecting the first object. Tab or Right Arrow accepts that inline text. Enter commits either an exact identity or a shorter query that still resolves to known suggestions; unmatched prose remains ordinary free-text search. The shared page-controls bar now follows that same owner split. frontend-modern/src/components/shared/PageControls.tsx stays the render shell for canonical page-level control composition, while frontend-modern/src/components/shared/FilterToolbar.tsx owns the shared search-row, filter-row, and inline-leading-slot layout surface. Monitoring pages that need workspace tabs or count chips next to search should route that through the shared searchLeading slot instead of recreating a second local header strip above the control bar.

Pages that filter a list-of-resources surface (Workloads, Storage, and the Proxmox Backups tab) compose the chip-based frontend-modern/src/components/shared/FilterBar/FilterBar.tsx shell instead of PageControls. Each page declares a FilterDef[] catalog (label, options, value, defaultValue, group); FilterBar renders chips for active filters and exposes the rest behind a "+ Filter" menu, with type-ahead at both the menu and chip popovers (AddFilterMenu and FilterChip). Low-frequency view options (grouping segmented control, columns picker, sort key) compose the shared ViewOptionsDisclosure through FilterBar's viewOptions prop instead of remaining as permanent toolbar controls. FilterBar owns the View trigger and inline disclosure; feature consumers pass only panel content and must not import or render ViewOptionsDisclosure themselves. Consumers with no currently applicable presentation choice must omit viewOptions rather than passing an empty conditional wrapper that leaves a dead View trigger. Contextual frequent actions may use leadingControls, while table counters, active trend ranges, and other persistent orientation readouts may use trailingControls. Platform tables inherit the same ownership through PlatformTableToolbar. View switches sit above the bar as navigation, not filters: the Proxmox Backups tab's By date / Coverage PlatformSectionTabs and the Storage subtabs (Pools / Physical Disks). FilterBar owns committed infrastructure search terms as removable pills, separate from its consumer-owned structured FilterDef chips. An exact infrastructure completion or a recognized abbreviated query clears the draft field, adds one search-term pill, and serializes committed terms as comma-separated inclusive alternatives so operators can select several objects without turning arbitrary phrases into chips. Removing a search-term pill must immediately update the consumer search state. Platform consumers project only canonical object identity, type, scope, status, and safe aliases through frontend-modern/src/features/platformPage/platformSearchSuggestions.ts; arbitrary resource metadata and secrets are not autocomplete candidates. PlatformTableToolbar must always expose contextual Clear filters for a non-empty search, including simple table-local search/status state that does not need a feature-owned reset. Route-owned or multi-facet platform surfaces may supply one composite reset, which the toolbar delegates exactly once; otherwise the shared FilterBar fallback clears its search and inline status catalog without forcing every platform table to repeat reset plumbing. The compact Add filter variant is the shared FilterBar default: the visible uppercase field label and labelled-field shell stay hidden while the native select retains its accessible Filter name. Platform table toolbars and other resource-list consumers inherit that treatment automatically instead of choosing different label and width chrome page by page. A surface that truly needs form-style field labelling must opt in explicitly through showAddFilterLabel. Primary filters with small, stable option sets should stay one-click controls inside that same FilterDef[] catalog by setting inline: true; FilterBar renders those as unlabeled compact segmented controls in the same second-row rail as view options, matching the v5 filter-bar pattern, and keeps longer or dynamic scope filters in the menu/chip path. Feature surfaces must not fork local filter rows or bury high-frequency Type, Status, or other true filter facets behind an extra menu just to regain one-click behavior; persistent presentation preferences such as grouped versus flat layout belong in View. Detailed multi-state catalogs are not primary one-click controls merely because they filter by status. A catalog large enough to create a horizontally clipped rail on narrow viewports, including Storage's seven-state status catalog, must stay in the Add filter menu and surface a non-default selection as a chip. Legacy PageControls and labelled select/toggle primitives are not the resource-list filter shape. If a future surface needs a new filtering affordance, it should extend the FilterBar catalog model or add a new registry-backed shared primitive rather than reintroducing a per-page select row.

Kiosk mode (?kiosk=1, the session flag read through frontend-modern/src/hooks/useKioskMode.ts) is the read-only wall display, and FilterBar owns hiding filter chrome there. While kiosk is on the shared bar stays unmounted, so search, inline and menu filters, View options, leading and trailing controls, and the search field's type-to-search listener are all absent on every consumer, including the Proxmox and vSphere workload overviews, platform tables through PlatformTableToolbar, Storage, Proxmox Backups, and Alert History. Filter state that is already applied keeps narrowing the rendered rows; kiosk changes what is shown, not what the URL or stored preferences select. Consumers must not re-implement the gate page by page: the Proxmox and vSphere overview toolbars lost it when they were hoisted out of WorkloadsSurface, and kept rendering on kiosk displays until the rule moved into the shared bar. A consumer whose surrounding chrome exists only for the bar, such as the Storage view subtabs, may still hide that chrome itself. The rule covers the shared bar only: a scope heading and its reset (the Proxmox Guests on <node> heading with Show all nodes), chart-driven selections on Proxmox Backups and Alert History, and search fields inside opened drawers are outside it.

The rest of the kiosk chrome belongs to the shared platform primitives in frontend-modern/src/features/platformPage/sharedPlatformPage.tsx, for the same reason. PlatformSectionTabs stays unmounted while kiosk is on, on every platform page and on the Proxmox Backups view rail: the URL picks the section and view the way it picks the platform, the maintainer's replies on #917 and #1102 promise a display without section navigation, Storage already hides its own subtabs there, and on a 400px rack LCD the rail cost 53px of height (106px on Backups, which carries two rails). createPlatformTablePreview shows every row in kiosk and renders no Show all control, so the Proxmox nodes table lists all of a 50-node estate instead of the first twelve. A wall display has nobody to press the button, and the primary navigation with its Alerts count is hidden there, so the table rows are where a node going down shows: on the mock estate the capped kiosk never showed the node that was offline. The expansion state is left alone, so leaving kiosk restores the rail and the preview as it was before kiosk (an expansion still resets once the rows no longer exceed the cap). A persisted expansion was rejected: the preview contract forbids a page-burying expanded default, and a display would still need one press per browser profile. Both rules sit in the shared primitives, so a page that adds a rail or a capped table must not add its own kiosk check. The controls the FilterBar rule leaves outside stay interactive on purpose. The Guests on <node> heading with its Show all nodes reset, and the chart selections on Proxmox Backups and Alert History, stay because a deliberate click on a row or chart, or a URL someone chose, can narrow the page and these are the display's way back from that narrowing, so hiding them would strand a touch display. Drawer search fields exist only inside a drawer someone opened with a click, and the Kubernetes Namespaces and Deployments and Docker Swarm drawers register type-to-search only while open, so useTypeToSearch stays kiosk-agnostic: with no such drawer open, a stray key on an overview reaches no field.

FilterBar does not carry a saved-views affordance. The former savedViewsKey / useSavedViews / SavedViewsMenu trio persisted named query strings to localStorage under pulse:filterbar:saved-views:<key>; it was removed because a saved view was only ever the page's URL query string, which the browser's own bookmarks already capture, sync, search, and share. The URL-ownership rule it depended on survives it and is now the primary contract: every filter a surface exposes must be URL-owned, so a filtered page is a shareable, bookmarkable link. Do not reintroduce an in-app view library; extend URL coverage instead. The Machines surface uses the canonical STANDALONE_QUERY_PARAMS query and status keys and delegates one composite reset to StandalonePageSurface; it must not mix a local search signal with a route-owned status facet or issue consecutive route writes that can resurrect one cleared parameter. The Proxmox Backups surface follows the same URL-ownership boundary with PROXMOX_BACKUPS_QUERY_PARAMS: search, workspace, scope, per-workspace facet, and selected activity day are all route-owned. Its feature owner may clear incompatible hidden facets when the workspace changes and may preserve that workspace during a composite reset, but the shared FilterBar remains the sole owner of Clear, filter-menu, and popover chrome across desktop and narrow layouts. ProxmoxBackupsCoverageStrip keeps its compact context on the title baseline at desktop widths and gives that context a full-width, left-aligned row below the title on narrow screens. A wrapped context must not retain an auto margin that creates an apparent empty column. That surface also keeps asynchronous posture presentation explicit: an unresolved canonical posture request uses a neutral Checking row state and verdict note, while host and orphan recovery rows without a canonical workload identity use neutral Not evaluated. Neither presentation state may reuse the server-owned Unknown label, enter the posture filter contract, or be styled as a protection success or failure. Alert History follows the same rule for its route-owned search, period, and severity state. Its feature hook owns one composite reset and exposes a search-aware active-state accessor to FilterBar; the shared shell owns where the contextual Clear filters action appears, while the alerts feature must not fall back to sequential per-control URL writes or hide that action for a search-only result set. Alert History severity option counts also follow the shared estate-orientation contract: the alerts feature supplies counts from the exact predicate used by its list, while FilterBar renders the values and the shared Inventory totals visibility preference decides whether they are shown. A feature must not build counts from page-wide alert totals after search or another active facet has narrowed the rendered rows, and a time-scope option must not claim a count from an unfetched period. The Alert History frequency axis is also a responsive contract. Desktop keeps the complete model-owned tick set; below the sm breakpoint the render surface keeps only start, midpoint, and end labels so locale-formatted timestamps do not overlap or create horizontal overflow. The 390px operator qualification must assert three visible labels, non-overlapping client rectangles, keyboard reachability, and a contained document before its actual-pixels receipt is recorded. The Alert History table's density tracks in index.css carry per-column widths through --alert-history-<column>-width custom properties set on the table at each density. Fixed layout sizes all nine tracks before collapsed ones are removed, so the visible widths sum to 100% and the table width is the sum of all nine; App.architecture.test.ts checks both at every density, the pixel minimums the content-bound tracks need at the narrowest card their density serves, that each density block sits inside its 50rem, 64rem or 80rem container query, and that Message keeps at least 18% of the card. Below a 50rem card the Alert History table's Resource and Message cells wrap in a line-clamp-2 wrapper that sets whitespace-normal itself (the global .table-fixed td rule forces nowrap), so more of a row reads without hover on a row that has no disclosure to recover the rest; Message keeps a two-line minimum height so its length never changes a row's height, which the shared windowing hook's single measured row assumes, and from a 50rem card both cells return to a one-line clamp (@min-[50rem]:line-clamp-1): icon-only Timeline/Resource row actions (100px instead of 177px) and a compact Status badge (108px instead of 119px) give Resource and Message the width that one line needs. A table row asks getAlertHistoryRowCopy for the subject-free copy, so Message leads with the alert's summary ("Offline") when it has one and keeps the full sentence as its title; cards keep the sentence. Alert History investigation detail is also responsive by interaction model, not only by CSS. Desktop may keep Timeline and Resource incident detail inline with its table row, while the virtualized phone card list must open the shared detail components inside the feature-owned full-height Dialog drawer in frontend-modern/src/features/alerts/MobileAlertHistoryInvestigationDialog.tsx. The drawer owns a bounded overflow-y-auto and overscroll-contain evidence scrollport, while the app scroll shell and the fixed-estimate history window remain unchanged underneath it. Escape and the explicit close action dismiss the drawer and restore focus to the originating row action when that virtualized row still exists, falling back to the phone history list when it does not. Future expandable content inside a fixed-estimate list must use the same independent-detail boundary or move to a variable-height virtualizer; it must not change a mounted virtual row's height and compensate with ad hoc page scroll writes. Because that popover combines view application, default selection, removal, and an inline naming form, it is a labelled non-modal dialog rather than an ARIA menu. Its trigger exposes the dialog relationship, Escape returns focus to the trigger (or from the naming form to the save action), the naming field has a persistent label, and destructive view controls remain visibly discoverable at narrow touch viewports instead of depending on hover. The panel anchors from the trigger's leading edge on the expanded mobile filter rail, where the trigger wraps to the left edge, and returns to trailing-edge alignment on desktop so its management actions cannot run past either viewport edge. Filter-bar popovers must compose the shared FilterPopoverTrigger; the shared trigger owns their matching text emphasis, icon geometry, disclosure chevron, active state, and button alignment instead of allowing each popup wrapper to restyle that contract independently. When no menu-backed filter row is present, contextual Clear filters joins the mobile action row instead of forcing View onto an isolated line. View takes the row's available trailing space and anchors its panel from the mobile action rail's trailing edge, then returns to trigger-relative trailing alignment on desktop; Saved keeps its leading-edge mobile anchor. When no menu-backed filter chip, contextual Clear action, or leading action is present, the compact Add filter control joins that same action cluster instead of occupying an otherwise empty row by itself. Once a menu-backed filter is active, Add filter remains with the active chips so scope editing stays grouped. The mobile action cluster is non-breaking as a unit; orientation readouts such as result counts and trend ranges wrap separately so they cannot strand View on a line by itself. The desktop controls rail keeps inline filters and utility actions at opposite edges while both fit on one line. Its wrapping parent owns that split through space distribution rather than an auto margin on the action cluster, so a wrapped Add filter / Saved / Clear / View row starts at the left edge instead of presenting an empty leading column. Feature surfaces must not compensate with page-local alignment or width overrides. The Add filter select and adjacent action/popover triggers share the canonical filterToolbarControlClass height. When the Add filter label is visually hidden, its FormSelect must also omit the labelled group's outer padded/ring shell so the native select does not become a double-framed, oversized control; the select uses one stable compact width instead of expanding to its longest hidden option. The visible Add filter value is a disabled, hidden placeholder, not an actionable option repeated at the top of the opened native list; only real filter values belong in that choice list. Counts or orientation strips presented as part of a filtered resource table must derive from that table's canonical filtered collection as well. They must not continue showing page-wide inventory totals after FilterBar state, saved views, or search has narrowed the rows the operator can see. Filtered summary strips should omit zero-value categories and their separators rather than turning absent states into persistent visual noise. Implicit "remember last filters" is intentionally not added — defaulting to yesterday's filter state on a monitoring page hides real problems. That same shared filter-toolbar boundary also owns controlled select continuity when filter options materialize asynchronously. LabeledFilterSelect must keep the caller-owned value visibly selected after option children arrive so dashboard, recovery, and other canonical filter bars do not drop their active selection until the operator reopens the control. The same primitive must keep its <label for> association reactive when a route-owned filter swaps the select id, label, and option set in place, so controls such as the workloads node/K8s cluster filter remain accessible after mode changes. That same boundary also owns live option propagation through shared page-control composition. Callers such as storage and recovery must pass source/filter option collections through reactive accessors instead of snapshot arrays when those options depend on post-load unified-resource state, so the shared toolbar can reconcile late-arriving options and preserved route selections without requiring page-local reset hacks. Shared default filter labels must also stay on the same primitive-level contract. Generic All … option text should route through frontend-modern/src/components/shared/filterOptionPresentation.ts, with domain presentation helpers supplying the noun phrase, so storage, alerts, recovery, settings, and future filter bars do not drift between title-case and sentence-case local strings. When those workspace tabs need an embedded control-bar treatment, they should still stay on the one canonical frontend-modern/src/components/shared/Subtabs.tsx primitive and reuse the established shell, list, and button class pattern already proven on owning surfaces like operations rather than introducing new variant APIs on the primitive. When that rail overflows on phone widths, Subtabs owns visible, accessible edge-scroll controls and keeps them in sync with native scrolling and rail resize; callers must not add drawer-local arrow overlays or leave clipped tab labels as the only overflow cue. Selection changes reveal the active tab by moving only that horizontal rail through the shared rail-visibility controller; Subtabs must not use scrollIntoView, which can also move page and drawer ancestors vertically. The search-input enhancement surfaces now follow that same owner split. frontend-modern/src/components/shared/SearchInputEnhancements.tsx stays the render shell, frontend-modern/src/components/shared/useSearchInputEnhancements.ts owns search-history persistence, menu-open lifecycle, blur commit policy, and tips/history interaction runtime, and frontend-modern/src/components/shared/searchInputEnhancementsModel.ts owns history-toggle copy plus history-menu button and row class policy. Future search-input-enhancement work should extend those owners instead of pushing history copy or menu presentation policy back into the shell. Search-history menus must remain full-width on narrow search surfaces while using a bounded desktop width aligned to the search field's leading edge; a full-page or full-toolbar search field must not turn the history popover into a screen-wide overlay that obscures unrelated controls. The shared search tips popover now follows that same owner split. frontend-modern/src/components/shared/SearchTipsPopover.tsx stays the render shell, frontend-modern/src/components/shared/useSearchTipsPopoverState.ts owns open-state, pointer/focus continuity, and outside-click/Escape listener runtime, and frontend-modern/src/components/shared/searchTipsPopoverModel.ts owns trigger variant, label/id defaults, hover policy, and trigger/popover class selection. Future search-tips work should extend those owners instead of pushing listener lifecycle or trigger policy back into the shared shell. Canonical customer disclosures inside shared shells route through frontend-modern/src/utils/docsLinks.ts, so settings privacy links resolve to shipped /docs/... assets instead of hard-coded GitHub main URLs that can drift from the running build. The pre-authenticated Login shell uses that same canonical docs-link boundary for self-hosted access recovery. TROUBLESHOOTING_DOC_URL routes the always- visible Can’t sign in? action to the shipped Troubleshooting guide, including when local login is hidden behind SSO, and the /docs/... public-route contract keeps that destination readable without an authenticated session. The login surface must remain guidance-only: it may not expose secrets, imply an email reset flow, or create a browser-side authentication bypass. The shared tooltip now follows that same owner split. frontend-modern/src/components/shared/Tooltip.tsx stays the render shell and singleton API boundary, frontend-modern/src/components/shared/useTooltipState.ts owns tooltip positioning lifecycle, RAF scheduling, and singleton visibility state, and frontend-modern/src/components/shared/tooltipModel.ts owns tooltip sanitization plus viewport-clamped positioning math. Future tooltip work should extend those owners instead of pushing singleton state, DOM measurement, or sanitization logic back into the shared shell. Shared portal-mounted tooltip shells such as frontend-modern/src/components/shared/TooltipPortal.tsx must use the same semantic surface tokens as the canonical tooltip instead of introducing light-mode-inverted palettes. That same tooltip owner now also holds the CSP-safe portal contract: shared tooltip shells must render through SVG/attribute positioning and viewport- clamped layout helpers rather than fixed inline left/top style attributes. When a shared portal tooltip is already visible, that same owner must reschedule positioning on live coordinate and viewport changes so chart hover tooltips keep following the active pointer instead of sticking to their first anchor. Only a visible tooltip holds a window resize listener. Hidden instances (one per metric bar or sparkline on every table row) and portals suppressed on a touch device register none and re-read the viewport when they open, so the listener count no longer grows with rows. TooltipPortal.test.tsx proves that hidden and touch-suppressed portals add no listener, that an open one re-clamps on a window resize, and that one opened after an unheard resize starts from the live viewport. Floating hover tooltips are a fine-pointer interaction only. The shared tooltip hook, portal, and singleton API must suppress them when the primary device reports no hover capability or a coarse pointer, because touch browsers may synthesize mouse-enter before the row click that should open canonical details. Desktop pointer and keyboard interaction remain unchanged; mobile row activation must continue directly to its drawer or other primary action. The shared collapsible search input now follows that same owner split. frontend-modern/src/components/shared/CollapsibleSearchInput.tsx stays the render shell, frontend-modern/src/components/shared/useCollapsibleSearchInputState.ts owns expand/collapse state, focus choreography, and type-to-search handoff, and frontend-modern/src/components/shared/collapsibleSearchInputModel.ts owns trigger-label, expanded-visibility, and full-width layout policy. Future collapsible-search work should extend those owners instead of pushing expand/collapse runtime or layout rules back into the shared shell. The shared pulse data grid now follows that same owner split. frontend-modern/src/components/shared/PulseDataGrid.tsx stays the render shell, frontend-modern/src/components/shared/usePulseDataGridState.ts owns breakpoint-driven min-width selection and stable-row reconciliation, and frontend-modern/src/components/shared/pulseDataGridModel.ts owns alignment class policy plus interactive-target row-click protection. Future pulse-data- grid work should extend those owners instead of pushing breakpoint lifecycle or interaction policy back into the shared shell.

The audit log settings surface now follows that same owner split. frontend-modern/src/components/Settings/AuditLogPanel.tsx stays the canonical SettingsPanel shell, while frontend-modern/src/components/Settings/useAuditLogPanelState.ts owns the license/paywall lifecycle, persisted filters, verification flow, and audit-log fetch orchestration. The shell must not re-accumulate localStorage or API runtime logic inline. Audit-log filter option labels must come from frontend-modern/src/utils/auditLogPresentation.ts and the shared filter-option label primitive instead of hard-coded title-case strings in the settings shell. When the shared runtime-capabilities store reports paid_runtime_required for audit_logging, Settings navigation must keep the Audit Log surface reachable and the panel must render paid-runtime-required copy with the private Pulse Pro download action. The runtime mismatch is not a plan upsell and must not be hidden by ordinary missing-feature navigation filtering. Audit-log fetch failures must preserve the structured backend error object through apiErrorFromResponse and render customer-facing copy from frontend-modern/src/utils/auditLogPresentation.ts; the settings shell may own refresh and pagination state, but it must not show raw Internal Server Error strings or unbounded page sizes as local hook behavior. Audit-log page loads are latest-request-wins. Changing the page size atomically resets the offset and starts one replacement request, and any superseded request is aborted or ignored so an older response cannot overwrite the new page. Page size is a durable table-presentation preference and belongs inside the shared FilterBar.viewOptions popover rather than as a permanent filter control. A failed load clears previously rendered events and totals, and a successful payload must contain an event array rather than treating null or an absent list as an empty audit history. That shared filter-option primitive is also the canonical owner for default All <scope> option wording wherever a product surface exposes filter selects or segmented filter choices. Workloads filters, storage source filters, Kubernetes namespace drawers, resource-change timeline filters, and alert configuration options must call frontend-modern/src/components/shared/filterOptionPresentation.ts through their nearest presentation/model owner instead of hard-coding page-local All ... labels.

The audit webhook settings surface now follows that same owner split. frontend-modern/src/components/Settings/AuditWebhookPanel.tsx stays the canonical SettingsPanel shell, while frontend-modern/src/components/Settings/useAuditWebhookPanelState.ts owns the license/paywall lifecycle, webhook fetch/save flow, validation, paywall tracking, and hidden-upgrade copy posture. The shell must not re-accumulate API calls or paywall tracking inline. Because it is the only webhook page in Settings, the shell opens with a signpost, in both the gated and enabled states, saying audit webhooks carry security events only and linking to the Alerts notifications tab for Discord, Slack, ntfy or email delivery (issue #1344 tried to add a Discord webhook here). The link is a plain anchor the app Router intercepts. The same paid-runtime-required route applies to Audit Webhooks: missing audit_logging caused by a community runtime must keep the panel reachable, hide normal upgrade-plan prompts, and present the private Pulse Pro runtime download action instead of describing the feature as an unlicensed Pro upsell.

The diagnostics settings surface now follows that same owner split. frontend-modern/src/components/Settings/DiagnosticsPanel.tsx stays the top-level diagnostics shell, while frontend-modern/src/components/Settings/useDiagnosticsPanelState.ts, frontend-modern/src/components/Settings/DiagnosticsResultsPanel.tsx, frontend-modern/src/components/Settings/diagnosticsModel.ts, and frontend-modern/src/utils/diagnosticsPresentation.ts own the diagnostics run/export lifecycle, results rendering, sanitization/model helpers, and customer-facing diagnostics copy. The shell must not re-accumulate inline API calls, export-download plumbing, diagnostics-card composition, or diagnostics surface copy. PBS diagnostics status badges render the backend's canonical monitored connected/state result. The one-off diagnostics request remains a nested probe result; when the probe and monitor disagree, the row must name both states instead of turning a successful live check into a green monitored badge or hiding recovery evidence behind an undifferentiated failure. That same diagnostics owner split also keeps maintainer analytics out of the customer diagnostics surface. DiagnosticsResultsPanel.tsx, diagnosticsModel.ts, and the diagnostics export path must not render or preserve commercial funnel, sales funnel, pricing/checkout conversion, or infrastructure onboarding telemetry from /api/diagnostics; those signals belong in admin-owned metrics surfaces instead of Settings support UI. frontend-modern/scripts/settings-diagnostics-boundary-audit.mjs, called by the canonical frontend audit runner, enforces that boundary by failing if the diagnostics API, diagnostics results panel, or diagnostics payload model reintroduce those analytics fields outside the defensive strip helper, and by failing if production customer frontend source reintroduces the retired commercial/onboarding analytics wrappers or /api/upgrade-metrics/events calls. That same audit also fails if the retired conversion/funnel or metering packages return under the compiled product licensing path, because Settings support diagnostics cannot be the only customer-facing guard if the normal product binary still carries the maintainer analytics pipeline. Diagnostics cards that summarize Docker and Podman agent coverage must use the shared docker source-platform label from frontend-modern/src/utils/sourcePlatforms.ts for their heading and body copy, so diagnostics results stay aligned with the governed settings/source-platform vocabulary instead of inventing a local runtime family label.

The settings shell registry now also treats extracted feature prop contracts as canonical shell inputs instead of reaching back into feature panels for type ownership. frontend-modern/src/components/Settings/useSettingsPanelRegistry.tsx must consume the direct Proxmox panel contract through frontend-modern/src/components/Settings/proxmoxSettingsModel.ts, so the registry stays a shell/composition owner and does not depend on ProxmoxSettingsPanel.tsx as though the panel still owned the runtime model.

The retired /operations route is unregistered rather than a compatibility redirect. Diagnostics, reports, and logs belong to the shared Settings shell instead of a bespoke page-local tab surface. Support-only navigation must therefore route through the shared settings owners rather than rebuilding a second route-level shell, and public demo posture must keep those support entries hidden from the Settings navigation instead of reviving a standalone operations page. that are unavailable in demo mode.

The dashboard overview route and its feature-owned summary surfaces are retired. Authenticated root entry now lands on the first visible provider/runtime platform, so first-viewport estate orientation belongs to that platform page plus the Add infrastructure flow rather than a separate dashboard or legacy Infrastructure shell. Future overview or brief-style surfaces must be governed as new product surfaces before they add route-level data orchestration, section anchors, or Assistant prompt handoffs; they must not restore frontend-modern/src/pages/Dashboard.tsx, frontend-modern/src/features/dashboardOverview/, or deleted dashboard-only presentation helpers as compatibility paths. The primary navigation active-tab contract follows that retirement boundary: retired or unknown routes such as /dashboard must not be coerced into the nearest platform tab just because the authenticated shell has a provider-first landing fallback. Shared desktop and mobile navigation must tolerate a missing active tab for those paths while still highlighting canonical active routes such as Proxmox, Docker, Kubernetes, TrueNAS, vSphere, Machines, Alerts, Patrol, and Settings. Recovery has no workspace switch of its own: the protected-items versus recovery-events Subtabs switch left with the aggregate Recovery page on 2026-05-26. The Proxmox Backups tab switches between /proxmox/backups/date and /proxmox/backups/coverage through the shared PlatformSectionTabs row above its FilterBar, and the TrueNAS Protection tab is a single table. A future recovery view switch belongs on those platform-page primitives, not on a recovery-local switcher, chip row, or filter-row embedding. The shared table primitives now also need to preserve caller-owned separator styling. TableHeader and TableBody may provide canonical default borders and dividers, but when a caller supplies explicit border or divide classes the shared primitive must defer to that local contract instead of silently forcing the default separator treatment back into the rendered DOM. Cell padding follows the same ownership rule. TableCell and TableHead provide the canonical px-2 sm:px-3 and py-* defaults, but Tailwind emits every padding utility at equal specificity in spacing-scale order, so a base utility silently beats any smaller caller utility. When a caller's class names a side's padding (p-*, px-*, pl-*, pr-*, py-*, pt-*, pb-*, with or without !), the primitive omits its default for that side and keeps it for the sides the caller left alone; prefixed-only variants such as lg:px-0 keep layering on the default. Tables therefore state the padding they want in their own classes or presentation constants instead of carrying !px-* overrides to beat the primitive. Table.test.tsx pins the base, single-side, important, p-*, prefixed-only and reactive cases. That same shared table boundary now owns CSP-safe fill rendering for metric bars: frontend-modern/src/components/shared/ProgressBar.tsx must render fill width through DOM attributes rather than inline width styles that break the public demo CSP. Recovery has no shell, filter controller, route serializer, or column model left for frontend primitives to govern. The aggregate Recovery page, its section owners under frontend-modern/src/components/Recovery/, and its state owner frontend-modern/src/features/recovery/useRecoverySurfaceState.ts were deleted on 2026-05-26 together with the rollup, facet, and series hooks (useRecoveryRollups.ts, useRecoveryPointsFacets.ts, useRecoveryPointsSeries.ts), and the recovery query serializer and recoveryItemTypePresentation.ts that only it read have been deleted since. Recovery rows render in two platform-owned tabs built from the shared platform-page primitives: the Proxmox Backups tab and the TrueNAS Protection tab. The Proxmox Backups tab parses /api/backups/pve and /api/backups/pbs inside ProxmoxBackupsTable.tsx and reads posture through frontend-modern/src/hooks/useProtectionPostures.ts. The TrueNAS Protection tab reads recovery points, and what binds that path is the decode boundary. frontend-modern/src/utils/recoveryPlatformModel.ts maps the transport names onto the canonical runtime model: provider to platform, subjectResourceId to itemResourceId, subjectRef to itemRef, and display.subjectLabel / display.subjectType to display.itemLabel / display.itemType. The points handler sends the item label only as display.subjectLabel, so that fold is what names TrueNAS Protection rows. frontend-modern/src/hooks/useRecoveryPoints.ts, the only recovery-points reader, applies normalizeRecoveryPointsResponse to every /api/recovery/points response. frontend-modern/src/types/recovery.ts declares the subject names on the transport types alone (RecoveryPointTransport, RecoveryPointDisplayTransport), so TrueNASProtectionTable.tsx and truenasPageModel.ts label, sort, and search rows by display.itemLabel, display.itemType, and itemRef, and a read of subjectRef or display.subjectLabel off a normalized point fails the type check. frontend-modern/src/api/__tests__/recoveryTransport.test.ts pins the type split and the decode that names the rows. New recovery filters, links, or view switches belong to the platform tab that owns the rows, on the platform-page primitives, rather than in a revived cross-platform recovery shell or a shared recovery serializer in frontend-modern/src/routing/resourceLinks.ts. The frontend decodes no rollups: the rollup normalizer and the rollup, series, and facet types went with their last readers. The retired dashboard action queue must not be reintroduced as a compact Patrol or infrastructure issue panel. Patrol-owned runtime findings remain governed by frontend-modern/src/utils/aiFindingPresentation.ts and their own Patrol route/store surfaces; any future cross-surface issue queue needs a new governed owner rather than reviving dashboard action-panel files.

Feature-owned alert shells under frontend-modern/src/features/alerts/ now also treat shared action runtime as a first-class feature owner instead of rebuilding it per surface. The overview shell must compose frontend-modern/src/features/alerts/useAlertAcknowledgementState.ts for acknowledge/restore behavior rather than keeping duplicate API and notification logic inline in useAlertOverviewState.ts or a revived dashboard recent-alert panel. That hook takes the shared alert store's required updateAlert and keeps no acknowledgement override: the store holds the optimistic state until the server confirms it, while a second copy in the hook outlived that confirmation and hid a later unacknowledge from another session until reload. Snooze state reads the same store alerts. The same feature-owner rule now applies to the alert scheduling surface: frontend-modern/src/features/alerts/tabs/ScheduleTab.tsx must remain the schedule render shell, while frontend-modern/src/features/alerts/useAlertScheduleState.ts owns schedule reset/update policy and canonical default application. The tab should not re-accumulate quiet-hours, cooldown, grouping, or escalation mutation logic inline. The thresholds editor now follows that same split more tightly: frontend-modern/src/features/alerts/thresholds/hooks/useThresholdsTableState.ts must stay the table-shell owner for route sync and local UI state, while frontend-modern/src/features/alerts/thresholds/hooks/useThresholdsData.ts stays the composition shell for threshold resource-family projectors, frontend-modern/src/features/alerts/thresholds/hooks/useThresholdsRecoveryDefaultsState.ts owns backup/snapshot default sanitization and factory-drift policy, and frontend-modern/src/features/alerts/thresholds/thresholdsOverrideMutationModel.ts owns pure override upsert/hysteresis/state-strip helpers, frontend-modern/src/features/alerts/thresholds/hooks/useThresholdsOverrideMutations.ts owns threshold-save and backup/snapshot override persistence, and frontend-modern/src/features/alerts/thresholds/hooks/useThresholdsAvailabilityMutations.ts owns availability-state policy and alert-removal side effects. The table-shell hook should not re-accumulate raw override mutation logic, recovery-threshold defaults policy, or resource-family projection engines inline. TrueNAS system threshold rows follow that same feature-owner split and the canonical alert identity chain. useThresholdsPlatformData.ts must project the current canonical resource ID as the writable storage ID, accept bounded superseded/metric-target candidates for legacy readback, and retain the projected row while editingId is active so WebSocket resource repolls, reordering, or changing display telemetry cannot reset the input. useThresholdsOverrideMutations.ts owns the blur commit: it removes every bounded candidate, writes exactly one current-ID raw override, marks the configuration dirty, and leaves the global Save Changes control to perform the API persistence. Multiple same-hostname TrueNAS systems must stay independent because configured connection identity, not display name or DMI serial, owns the row. The regression boundary is frontend-modern/src/features/alerts/thresholds/hooks/__tests__/truenasThresholdPersistence.test.tsx; the global payload boundary is frontend-modern/src/features/alerts/__tests__/useAlertsConfigurationState.test.tsx.

The updates settings surface now follows the same presentation-owner rule. Source builds have no published release-update target. The shared update store must discard a cached release offer when the current runtime reports isSourceBuild or isDevelopment, even if its version string did not change. The update panel labels this state Source build, disables release checks and automatic updates, and omits release notes. UpdateInstallGuide suppresses release install actions and generic Docker pull commands for source builds, including stale cached offers. Docker source builds instead explain manual image replacement. Stable and preview release flows retain their existing checks and installation guidance. The update-store, install-guide, and presentation tests own this cross-control contract. frontend-modern/src/components/Settings/UpdatesSettingsPanel.tsx stays the top-level settings shell, while frontend-modern/src/components/Settings/UpdateInstallGuide.tsx, frontend-modern/src/components/Settings/CopyCommandBlock.tsx, and frontend-modern/src/components/Settings/updatesSettingsModel.ts plus frontend-modern/src/utils/updatesPresentation.ts own the deployment-specific install guide, copy-command block, and update-channel/install model data plus customer-facing update status/action copy. frontend-modern/src/components/Settings/UpdateHistorySection.tsx joins that split as the presentation owner for the update history table and the rollback confirmation dialog: the panel shell mounts it as a section and must not inline history rows, rollback gating, or rollback confirmation copy itself, and the section starts rollbacks through the shared updateStore.rollbackUpdate action rather than its own POST path. The panel shell must not rebuild copy-to-clipboard command cards, deployment instruction trees, or update-surface wording inline. CopyCommandBlock must use the shared copyToClipboard helper so install/update/agent snippets keep the same Clipboard API fallback path and only report copied state after the shared copy path succeeds. The running published version must keep a direct Current release notes link in this settings shell so suppressing or dismissing the one-time post-update notice never makes the changelog undiscoverable. Development and source-build identities must not render that link as if they named a published release.

The update verdict is honest about its age. The frontend serves the update verdict from a 24-hour localStorage cache, so the panel's "Up to date" state must render the age of the check it came from rather than reading as a live comparison (#1601). frontend-modern/src/stores/updates.ts exposes lastCheckedAt, the epoch milliseconds of the check backing the currently displayed verdict, set on fresh checks, on cached-verdict reuse, and on the cached fallback after a failed check. frontend-modern/src/utils/updatesPresentation.ts owns the customer-facing wording through getUpdateCheckedLabel ("Checked 3 hours ago", "Not checked yet"); the panel shell must not inline that copy and hides the line for source builds, where update checks are disabled. Regression boundary: frontend-modern/src/components/Settings/__tests__/settingsArchitecture.test.ts.

The reporting operations surface now follows the same shell-state-model rule. frontend-modern/src/components/Settings/ReportingPanel.tsx stays the operations-panel shell, while frontend-modern/src/components/Settings/useReportingPanelState.ts owns the license/trial lifecycle and report generation flow, frontend-modern/src/components/Settings/reportingPanelModel.ts plus frontend-modern/src/utils/reportingResourceTypes.ts own the request/range/filename model and reporting-type API mapping, frontend-modern/src/components/Settings/ResourcePicker.tsx plus frontend-modern/src/utils/reportableResourceTypes.ts own the reportable resource selection, filter, sort, and empty-state contract, and frontend-modern/src/utils/reportingPresentation.ts owns the user-facing range/status copy. Consumers import @/utils/reportingResourceTypes directly; the former one-line compatibility re-export under components/Settings/ was removed as dead code once its last importer moved. Native Kubernetes inventory-only resource types, including ReplicaSets, EndpointSlices, NetworkPolicies, StorageClasses, ConfigMaps, Secrets, ServiceAccounts, Roles, ClusterRoles, RoleBindings, ClusterRoleBindings, ResourceQuotas, LimitRanges, PodDisruptionBudgets, and HorizontalPodAutoscalers, must map to the existing reporting k8s transport token at this edge rather than widening the metric-report picker into platform-object inventory. Kubernetes RBAC inventory (Roles, ClusterRoles, RoleBindings, ClusterRoleBindings) joins the existing K8s inventory bucket on that transport without introducing a separate reporting token: from the reporting transport's point of view it is platform-object inventory the same way ConfigMaps and ServiceAccounts already are, even though the rendered Configuration tab surfaces it through RBAC-specific lifecycle/data-shape columns inside KubernetesConfigTable. The shell must not re-accumulate license bootstrapping, inline report API requests, blob-download plumbing, or local resource-type filter and reporting-token maps.

General settings segmented selectors for theme preference and temperature unit must now also route through the shared FilterButtonGroup primitive instead of maintaining local button-group styling forks inside frontend-modern/src/components/Settings/GeneralSettingsPanel.tsx.

Reporting time-range/export selectors and General settings Proxmox VE polling presets must now also route through the shared FilterButtonGroup prominent variant instead of maintaining local blue segmented-control styling forks in feature components. That same shared FilterButtonGroup primitive must stay CSP-safe: touch-scroll overflow behavior must come from canonical CSS classes rather than inline style attributes so settings and reporting selectors do not reintroduce browser console CSP violations under the release build policy.

Selectable settings cards for compact provider pickers and detail choice panels must now route through the shared SelectionCardGroup primitive instead of duplicating border-2 active-card styling in feature components.

Settings informational, warning, success, and danger callouts with icon-plus-copy layouts must now route through the shared CalloutCard primitive instead of maintaining feature-local colored bordered wrappers. The primitive owns the tone palette and the scale="compact" density used by smaller settings notices, and the settings-callout-card-shell shared-template registry rule requires current settings consumers to compose it instead of reintroducing local panel shells. Connection-editor status, feature-disabled, delete-error, and probe-result notices are part of that same settings callout boundary: the editor and credential slots own the source-specific lifecycle or API meaning, while CalloutCard owns the warning/success/danger shell and compact density. The settings-connection-editor-local-*-callout-shell pattern guards block future connection-editor files from reintroducing amber, red, or rose local notice shells. Shared error-boundary fallbacks use the same boundary: the fallback owns error copy and reset/reload handlers, while CalloutCard owns danger tone, spacing, dark-mode styling, and alert layout instead of inline red panels or raw SVG alert glyphs. Update confirmation and progress modals use the same shared boundary: the modal flow owns update state and copy, while CalloutCard, Button, ActionIconButton, LoadingSpinner, and lucide icons own the colored notice, command, icon-only close, and status indicator chrome.

Settings loading placeholders must route through the shared SettingsLoadingSkeleton primitive instead of local animate-pulse blocks. Feature panels may choose the loading shape and row counts, but the shared primitive owns pulse animation, skeleton fill tokens, metric-card grids, progress-card rows, table header/body shells, and labelled status semantics. The settings-loading-skeleton-shell registry rule covers the current organization, security overview, and resource data policy loading surfaces, the settings-loading-state-shared-skeleton-required guard requires future Settings *LoadingState files to compose that primitive, and the settings-local-loading-skeleton-block-shell guard blocks local pulse skeleton blocks from returning inside Settings components.

Settings external documentation text links must route through ExternalTextLink, while button-styled external actions route through ButtonLink/UpgradeButtonLink. Shared primitives own new-tab safety, rel policy, link tone, compact action density, and focus styling; settings panels must not hand-code raw <a target="_blank"> anchors for documentation links. The settings-external-text-link-shell and settings-external-text-link-local-anchor registry entries enforce that split, and the Button registry owns the info variant for blue documentation CTAs. A documentation link inside a tinted callout sentence (the backup export and import warnings) uses the inlineInherit variant, which keeps the callout's text colour and underline instead of switching to link blue.

Platform inline notices that sit inside platform pages but are not settings callouts must route through the shared InlineNotice primitive. Platform owners provide only the affected-resource copy, render predicate, and destination; the shared primitive owns the dense warning/info/danger/success tone palette, icon/content layout, and action-link chrome. The platform-inline-notice-shell registry rule covers current outdated-agent and outdated-sensor notices, and the platform-inline-notice-local-amber-shell pattern guard blocks future platformPage files from reintroducing page-local amber notice shells. A notice's render predicate reads what its page hydrates on every tab the notice renders on. collectOutdatedSensorSetupNodes lists the PVE nodes that carry the registry's proxmox.sensorSetupOutdated verdict and takes no disk rows, because only the Proxmox Storage tab loads them; a predicate over them left the notice silent on Overview. A notice names each affected resource by the label its page's table shows for the row. The sensor-setup notice leads with the node's trimmed resource name (the operator-set display name when one exists, else the Proxmox node name), falling back to the id as the nodes table does, and adds the Proxmox node name in parentheses unless the label already contains it as a whole word, ignoring ASCII case, or as the host part of a domain name, as in West Production C (pve3). Both names are needed: the nodes table switches to the Proxmox node name in its narrowest layout, and the setup script is re-run on that Proxmox node, so the rule is stricter than hasAlternateDisplayName in the guest table's node group header and keeps pve-3 beside a PVE3 label.

Alert incident-event filter containers, labels, and chips must now route through the shared presentation helpers in frontend-modern/src/utils/alertIncidentPresentation.ts instead of allowing frontend-modern/src/pages/Alerts.tsx and frontend-modern/src/features/alerts/OverviewTab.tsx to fork their own filter button styling.

Alert incident acknowledged badges, event cards, and note-editor controls must also route through frontend-modern/src/utils/alertIncidentPresentation.ts instead of letting the alerts page and overview timeline maintain duplicate inline incident-detail styling.

Alert incident meta-row and detail-text presentation must also route through frontend-modern/src/utils/alertIncidentPresentation.ts instead of letting the alerts page and overview timeline maintain duplicated inline incident typography rules.

Alert incident timeline event card structure must also route through frontend-modern/src/components/Alerts/IncidentTimelineEventCard.tsx so the alerts page and overview timeline share one canonical event-card renderer instead of reimplementing the same summary/detail/output block twice.

The full expanded alert incident detail panel and event-filter controls must also route through frontend-modern/src/components/Alerts/IncidentTimelinePanel.tsx and frontend-modern/src/components/Alerts/IncidentEventFilters.tsx rather than rebuilding loading/error copy, filter controls, note-editor wiring, or event-card composition separately inside the alerts page and overview tab.

Resource incident panel card and summary-row presentation must also route through frontend-modern/src/utils/alertIncidentPresentation.ts instead of maintaining page-local incident panel styling inside frontend-modern/src/pages/Alerts.tsx.

The settings shell now also has an explicit five-way ownership split. frontend-modern/src/components/Settings/useDiscoverySettingsState.ts owns the shared discovery draft and subnet-validation state, frontend-modern/src/components/Settings/useSettingsInfrastructurePanelProps.ts owns infrastructure workspace prop assembly and resource-derived infrastructure read-model shaping for the shell, frontend-modern/src/components/Settings/settingsNavigationModel.ts owns settings tab identity, canonical route derivation, route eligibility, and retired infrastructure/workloads alias rejection. settingsRouting.ts and settingsTypes.ts remain thin compatibility re-export shims only, so external consumers can bridge to the canonical owner without reintroducing a second settings navigation model. settingsNavCatalog.ts owns settings navigation metadata and item lookup, settingsNavVisibility.ts owns feature/capability visibility and lock policy for settings navigation, useSettingsNavigation.ts owns reactive URL sync and canonical tab-selection state, SettingsDialogs.tsx owns shared settings modal composition, including the route-owned billing focus contract where /settings/system/billing/plan is the canonical settings-tab destination, /settings/system/billing/usage is a same-tab child state, and legacy billing base/hash links are compatibility inputs rather than primary runtime routes. Infrastructure settings no longer has route-level platform-selection state: /settings/infrastructure and /settings/infrastructure?add=<step> are the only routeable Infrastructure settings entry points for platform/API and agent-backed source setup. Agentless ping/TCP/HTTP checks are monitoring availability settings at /settings/monitoring/availability, with /settings/monitoring/availability?add=target as the route-owned add dialog. Machine entry points must add targetKind=machine, while the focused availability checks entry points for services and devices must add targetKind=service, so deep links open the same owned dialog with the correct bounded target kind already selected. That target kind only scopes the availability form copy and payload; it must not make agentless reachability targets eligible for the Machines table. Former nested aliases such as /settings/infrastructure/install, /settings/infrastructure/platforms/proxmox/pbs, /settings/infrastructure/api/pve, and /settings/workloads/docker must fail route eligibility instead of being normalized back into the Infrastructure workspace. That same settings access boundary must keep route eligibility separate from sidebar visibility. Panel-owned feature gates such as Relay, Reporting, RBAC, Audit Log, and Audit Webhooks may be hidden from the navigation on Community installs, but their direct settings routes must stay routeable so the owning panel can render its locked, non-flashing state instead of being bounced to the default Infrastructure tab. useSettingsShellState.ts owns shell-local sidebar/search/password-modal state, and settingsTabSaveBehavior.ts owns settings tab save-behavior lookup, frontend-modern/src/components/Settings/useSettingsSystemPanels.tsx owns system panel prop assembly for general, network, updates, and recovery, and frontend-modern/src/components/Settings/settingsPanelRegistryContext.tsx owns registry context assembly for dispatchable settings tabs while frontend-modern/src/components/Settings/settingsPanelRegistryLoaders.ts owns the lazy settings panel loader table and route-to-panel import boundary, and frontend-modern/src/components/Settings/useSettingsPanelRegistry.tsx owns the final memoized registry composition only. frontend-modern/src/components/Settings/Settings.tsx must stay a shell that wires those owners together instead of re-accumulating infrastructure workspace props, registry context maps, system panel prop maps, lazy loader definitions, or discovery draft state inline. That same settings-routing contract now also owns the Support group for Diagnostics & Health, Data & Reports, and System Logs: the navigation model must reject old /settings/operations/* settings paths instead of normalizing them into /settings/support/*, and the top-level /operations/* browser path must stay unregistered. The catalog plus visibility owners must still treat support surfaces as Settings-native pages rather than as a second top-level utility destination.

The resource incident panel's collapsed activity summary is now part of that same shared primitive boundary. Event-type count chips, visible-event copy, and the summary-ordering helper in frontend-modern/src/features/alerts/types.ts must stay shared across alert timeline surfaces instead of rebuilding page-local event summaries or bespoke incident-card markup.

Feature-owned route surfaces under frontend-modern/src/features/ must also keep their shell/runtime split explicit once a subsystem grows real transport or polling lifecycle. The Patrol feature is the current reference shape: frontend-modern/src/features/patrol/PatrolIntelligenceSurface.tsx stays the feature shell, reads its active workspace view from the route through patrolWorkspaceRouting.ts rather than holding it in component state, frontend-modern/src/features/patrol/usePatrolIntelligenceState.ts owns the runtime state machine, frontend-modern/src/features/patrol/patrolInvestigationContextModel.ts owns the pure Patrol-to-Assistant handoff and operator briefing derivation, frontend-modern/src/stores/aiIntelligenceSummaryModel.ts owns canonical AI summary normalization at the shared store boundary, and the Patrol-owned header, banner, attention workbench, workspace, and panel section files under frontend-modern/src/features/patrol/ own the heavy render surfaces. Shared shell governance should reinforce that pattern instead of letting feature render surfaces re-accumulate API and timer orchestration inline. That same route-owned page-health rule now also applies to Patrol: a feature surface may not present a green or all-clear primary summary when the owning runtime contract says the page is blocked or unavailable, even if the last successful snapshot was healthy. That same rule also applies to compact Patrol summary fragments inside the feature surface: count-only strips or metric cards must not emit No issues found or other reassuring copy when the owning overall-health summary is degraded or not fully verified. Patrol renders no summary shell today; the header recency line shows the latest check or activity time, how many resources that check covered, and the next scheduled check. If a summary shell returns, it should also surface verification scope from the owning run-history contract. Operators should be able to see, inside the same summary surface, whether Patrol recently completed a full verification pass or whether recent activity was limited to scoped/erroring patrol runs. When the same governed run-history contract shows a recent full patrol plus same-day scoped follow-up work, that summary shell should also carry a compact activity-mix explanation rather than forcing operators to infer why Patrol looked busy from a second competing status band. That explanation belongs on the verification surface itself when operators are reconciling Recently checked copy against same-day scoped Patrol bursts; the supporting activity context may complement the readout, but it is not sufficient as the only explanation path. Patrol recency labels and coverage phrases come from the Patrol recency presenter (getPatrolRecencyPresentation), never from hardcoded shell copy. The header labels the latest completed run of any type Last check; without run history it uses Last check for the last_patrol_at full-sweep transport fact and Last activity when last_activity_at is newer, and it never collapses them into a generic Last run label. Its coverage phrase reads checked N resources and never says verified. The current-findings empty state already applies that run-history rule through getPatrolRunCoverage: a successful latest full patrol with non-zero resource coverage, completed within the last 24 hours with no failed run after it, keeps an overall-health coverage factor from turning the empty queue into Check needed or Patrol needs review. If a summary shell returns, it must not present Recent coverage is incomplete in that state either. That same Patrol shell ownership includes refresh affordance state: frontend-modern/src/features/patrol/usePatrolIntelligenceState.ts must keep operator refresh controls generation-aware, timeout-bounded, and separate from background polling state, so a slow supporting intelligence read cannot make the Patrol stale-data Retry action spin indefinitely or stay disabled while Patrol findings and status remain visible. Neither the Patrol load and poll path (the store's loadDashboardData bundle) nor the Retry path's background supporting reads fetch the global learned-correlation list or the provider circuit-breaker status, neither of which any Patrol surface shows. That same Patrol shell should make scoped trigger policy legible without another navigation step. frontend-modern/src/features/patrol/PatrolIntelligenceHeader.tsx should keep actionable scoped-trigger state legible without promoting background-only policy pauses as default operator guidance; run-trigger details belong in run history or explicit secondary context. That same Patrol-facing primitive vocabulary must stay product-first. Patrol summary actions, runtime banners, run-history runtime-failure actions, runtime-finding actions, circuit-breaker copy, and Patrol control/provider controls may point at the shared provider settings route or model catalog, but they should describe those controls as Patrol/provider surfaces through frontend-modern/src/utils/patrolRuntimeActions.ts rather than falling back to generic AI Settings, AI Model, or AI circuit breaker copy inside the Patrol shell itself. That same product-first naming rule also applies to Pulse Intelligence settings: frontend-modern/src/components/Settings/AISettings.tsx, frontend-modern/src/components/Settings/settingsHeaderMeta.ts, frontend-modern/src/components/Settings/settingsNavCatalog.ts, frontend-modern/src/components/Settings/useAISettingsState.ts, and frontend-modern/src/utils/aiSettingsPresentation.ts must present the provider surface to operators under the Pulse Intelligence settings group as Provider & Models provider/model configuration rather than as a generic AI Services shell. The canonical browser route for that surface is /settings/pulse-intelligence/provider; legacy /settings/system-ai links may remain routeable compatibility aliases, but new navigation, setup, Assistant, and Patrol repair CTAs must emit the Pulse Intelligence route. On the main Patrol page, though, governed activity context belongs inside frontend-modern/src/features/patrol/PatrolIntelligenceWorkspace.tsx as current work, selected run history, or explicit details. Do not reintroduce a parallel page-level status strip above the current-work workspace. That same composition rule applies to the workspace: the default path should move directly into findings and run history instead of repeating runtime context through a second pre-tab status strip. Supporting context follows that same composition rule. Recent changes, learned correlations, and policy coverage stay off the Patrol workspace: it has no Details control or supporting-context panel for them in any state, with or without active findings or a selected run, so that evidence never competes with current work as a peer workflow or reads as a fresh Patrol result. Selected-run history should also suppress generic findings filter chrome and read as a Patrol run record. Missing legacy finding_ids remains an internal fail-closed scoping condition, but the visible caveat should say the finding record is unavailable rather than exposing snapshot/filter vocabulary. Workspace section descriptions must use Patrol-owned mode presentation copy that reflects the selected mode and lock state; they must not hardcode an all-mode sentence that tells watch-only users Patrol can investigate, ask for approval, or fix issues. The Patrol workspace must not add a generic Details panel to explain learned correlations, recent changes, or policy buckets; those signals may support Assistant and backend reasoning without becoming default operator chrome. Setup-only Patrol runtime failures must use the Patrol-owned Fix Patrol setup workspace title and setup description plus a dedicated setup task with the direct provider-settings action, rather than presenting that state as a normal Current issues infrastructure queue, and they must suppress generic issue-row chips, expand chevrons, and Active / All / Resolved filter chrome in that setup-only state. That setup-only workspace must also be the only visible provider-repair CTA for that state: it should use the Open Provider & Models action and suppress the readiness banner so setup does not appear twice. Run history must stay hidden as a competing action until Patrol can check infrastructure or an operator is already reviewing a specific run record.

Shared primitive consumers that split status-dot tone and status-text tone must now keep both values routed through the same exported presentation helper. Feature cards such as RAID status may not call shadow local aliases that drift from the canonical shared class/variant helpers.

Alert resource display labels used by the thresholds editor and alerts page must now route through the shared helper in frontend-modern/src/features/alerts/helpers.ts instead of rebuilding resource display-name fallback chains inline. Governed resources must preserve their canonical policy-aware label across grouped node headers, docker host grouping, and saved override rows rather than collapsing back to raw names or friendly-name truncation.

Shared search inputs must now keep their forwarded keyboard, blur, and clear handlers as explicit callable functions instead of relying on loose Solid event-handler unions. Shared search primitives still need to accept the real input/button event targets, but direct invocation inside the primitive must stay type-safe so consumers do not reintroduce union-call regressions while adding history, shortcut, or trailing-control behavior.

Shared shared-shell primitives that expose semantic title or value-level onChange props must now explicitly omit the conflicting DOM attribute names from their inherited HTML props. CalloutCard, FilterSegmentedControl, and Subtabs may still forward ordinary div attributes, but their canonical API must preserve JSX element titles and value-callback handlers instead of widening back to raw DOM attribute unions.

Shared entitlement/migration warning banners that live under frontend-modern/src/components/shared/ must also keep their counted fleet surface on the Pulse Unified Agent term. Shared primitive copy may describe legacy/API-connected resources separately, but it may not regress the primary banner label or CTA text back to host-agent product language. The self-hosted commercial paywall copy on those shared warning surfaces is now also explicitly locked to monitored systems rather than agents. When a shared banner or shared settings shell would explain monitored-system plan caps, the correct primitive decision is absence: monitored-system volume is not a current paid-capacity surface. Shared headers and descriptions may use monitored-system language for inventory grouping and support ledgers, but they must not talk about monitored-system limits, cap pressure, plan capacity, admission freezes, or upgrade actions. Future work must not recreate MonitoredSystemLimitWarningBanner, its state hook, or banner-local monitored-system copy strings. Shared frontend label-formatting helpers now also have an explicit owner here. frontend-modern/src/utils/textPresentation.ts is the canonical shared owner for token humanization, identifier label formatting, title-casing, and arrow-delimited label presentation used across AI, Patrol, Storage/Recovery, and other feature surfaces. Feature contracts may depend on that helper, but they should not re-home or fork those generic text-formatting rules into feature-local utilities. That same shared presentation boundary now also owns operator feedback and shared table-label semantics. frontend-modern/src/components/Toast/Toast.tsx stays the render shell for the global toast stack, frontend-modern/src/utils/toast.ts owns the app-level trigger helper, frontend-modern/src/utils/semanticTonePresentation.ts owns canonical toast and diagnostics tone classes, frontend-modern/src/utils/emptyStatePresentation.ts owns the shared empty-state tone styling consumed by EmptyState, and frontend-modern/src/utils/typeColumnPresentation.ts owns the single canonical type-column label used across dashboard and alert tables. Future feedback, empty-state, or shared type-column work should extend those helpers instead of reintroducing panel-local tone classes, app-local toast wiring, or copy drift between tables. First-session educational surfaces must also stay brief, flat, and model-led. When Pulse needs to teach a user how a flow works, the primary on-screen guidance should collapse to a few short descriptions of the real product mental model instead of a logo wall, feature brochure, or verbose internal mechanics dump. The runtime wizard itself now stays on the two-step Welcome -> Security path, while the separate setup-completion preview owns the brief three-step explanation: install the Unified Agent, get the first Pulse resource, then layer on additional context.

The settings shell is now also a governed frontend primitive boundary. frontend-modern/src/utils/settingsShellPresentation.ts now owns the customer-facing settings-shell framing copy for navigation, search, loading, and unsaved-change banners so SettingsPageShell.tsx stays a render shell instead of re-accumulating product wording inline. At phone widths that shell is a two-level preference workspace rather than a compressed desktop card: the searchable grouped Settings index owns level one, the sticky back/title bar owns level two, and content uses edge-to-edge shallow section framing with 44-pixel interactive targets. Preference labels and their current controls should remain in one scan row where they fit; verbose trust, environment, and explanatory copy must be subordinated through a short summary, bounded line clamp, or explicit detail/documentation action instead of consuming the default viewport. Desktop retains the persistent sidebar, page description, and roomier panel spacing. Future top-level Settings work must extend SettingsPageShell.tsx and SettingsPanel.tsx rather than recreating a second mobile shell or returning to nested desktop padding.

The alerts page shell now follows that same page-shell rule for feature tabs: frontend-modern/src/pages/Alerts.tsx owns navigation and cross-surface routing, while feature-owned tab surfaces such as frontend-modern/src/features/alerts/tabs/DestinationsTab.tsx and frontend-modern/src/features/alerts/tabs/HistoryTab.tsx plus frontend-modern/src/features/alerts/tabs/ScheduleTab.tsx and frontend-modern/src/features/alerts/tabs/ThresholdsTab.tsx own their tab-local rendering and interaction logic. Future alert tab cleanup should continue by extracting page-local tab blocks into feature modules rather than expanding the top-level page file again, and history-table behavior or thresholds-table adapter logic should stay feature-owned unless it graduates into a shared primitive used by more than one alert surface. Within that thresholds surface, frontend-modern/src/components/Alerts/ThresholdsTable.tsx is now explicitly a shell consumer rather than the data or controller owner, and the tab render owners live in frontend-modern/src/components/Alerts/ThresholdsTableProxmoxTab.tsx, frontend-modern/src/components/Alerts/ThresholdsTablePMGTab.tsx, frontend-modern/src/components/Alerts/ThresholdsTableAgentsTab.tsx, frontend-modern/src/components/Alerts/ThresholdsTableDockerTab.tsx, frontend-modern/src/components/Alerts/ThresholdsTableKubernetesTab.tsx, frontend-modern/src/components/Alerts/ThresholdsTableTrueNASTab.tsx, frontend-modern/src/components/Alerts/ThresholdsTableVMwareTab.tsx, and frontend-modern/src/components/Alerts/ThresholdsTablePBSTab.tsx. frontend-modern/src/features/alerts/thresholds/hooks/useThresholdsTableState.ts owns the platform-shaped thresholds sub-route contract: /alerts/thresholds/proxmox, /alerts/thresholds/docker, /alerts/thresholds/kubernetes, /alerts/thresholds/truenas, /alerts/thresholds/vmware, /alerts/thresholds/pbs, /alerts/thresholds/pmg, and /alerts/thresholds/systems. Legacy neutral links such as /alerts/thresholds/infrastructure, /alerts/thresholds/containers, and /alerts/thresholds/mail-gateway must redirect to the matching platform-shaped route; legacy /alerts/thresholds/agents links must continue to resolve to Systems. Thresholds section state is also owned there and composed through the shared collapsible-section pattern. Resource sections open collapsed by default on desktop and narrow layouts, while persisted operator choices, Expand all, and Collapse all stay authoritative. The thresholds shell must pair that compact default with an immediately visible custom-override summary whose actions open the selected section under the Custom-only filter, so inherited state remains clear without forcing every table open. Metric editors on desktop, mobile, and bulk surfaces use explicit On/Off controls and positive enabled-value bounds; legacy values at or below zero still read as Off, while the backend disable sentinel remains compatibility data rather than customer-facing input or copy. The Proxmox tab is itself now a shell that composes frontend-modern/src/components/Alerts/ThresholdsTableProxmoxNodesSection.tsx, frontend-modern/src/components/Alerts/ThresholdsTableProxmoxPBSSection.tsx, frontend-modern/src/components/Alerts/ThresholdsTableProxmoxGuestsSection.tsx, frontend-modern/src/components/Alerts/ThresholdsTableProxmoxGuestFilteringSection.tsx, frontend-modern/src/components/Alerts/ThresholdsTableProxmoxBackupsSection.tsx, frontend-modern/src/components/Alerts/ThresholdsTableProxmoxSnapshotsSection.tsx, and frontend-modern/src/components/Alerts/ThresholdsTableProxmoxStorageSection.tsx using the shared contract in frontend-modern/src/features/alerts/thresholds/thresholdsTableSectionProps.ts. Future infrastructure-thresholds presentation changes should extend those section surfaces rather than restoring mixed JSX ownership to frontend-modern/src/components/Alerts/ThresholdsTableProxmoxTab.tsx. The Proxmox threshold tab includes a separate Guest Filesystems section for QEMU guest-agent mounts. Rows use the live per-filesystem alert resource ID for status and delay actions, but carry explicit override candidates and a stable override storage ID into the shared mutation owners. This split is required: using the alert ID as the persistence key strands settings on VM node moves, while using only the persistence ID breaks active-alert and intent-policy linkage in the shared resource table. The Docker tab now follows that same composition pattern through frontend-modern/src/components/Alerts/ThresholdsTableDockerIgnoredPrefixesSection.tsx, frontend-modern/src/components/Alerts/ThresholdsTableDockerServiceGapSection.tsx, frontend-modern/src/components/Alerts/ThresholdsTableDockerHostsSection.tsx, and frontend-modern/src/components/Alerts/ThresholdsTableDockerContainersSection.tsx. Future Docker thresholds presentation changes should extend those section surfaces rather than restoring mixed JSX ownership to frontend-modern/src/components/Alerts/ThresholdsTableDockerTab.tsx. The systems tab now follows that same composition pattern through frontend-modern/src/components/Alerts/ThresholdsTableAgentsResourcesSection.tsx and frontend-modern/src/components/Alerts/ThresholdsTableAgentDisksSection.tsx. Future systems-thresholds presentation changes should extend those section surfaces rather than restoring mixed JSX ownership to frontend-modern/src/components/Alerts/ThresholdsTableAgentsTab.tsx. The thresholds tab adapter contract now lives in frontend-modern/src/features/alerts/thresholds/thresholdsTabModel.ts, so frontend-modern/src/features/alerts/tabs/ThresholdsTab.tsx stays a thin shell instead of carrying a duplicate table adapter contract inline. That adapter must bridge function-valued selectors and mutation props into frontend-modern/src/components/Alerts/ThresholdsTable.tsx explicitly; spread- based table prop adapters are not allowed here because they can collapse function props on the live Solid surface and break thresholds runtime state. Canonical threshold row shaping now routes through frontend-modern/src/features/alerts/thresholds/thresholdsResourceModel.ts plus the family-owned feature hooks frontend-modern/src/features/alerts/thresholds/hooks/useThresholdsHostData.ts, frontend-modern/src/features/alerts/thresholds/hooks/useThresholdsDockerData.ts, frontend-modern/src/features/alerts/thresholds/hooks/useThresholdsGuestData.ts, frontend-modern/src/features/alerts/thresholds/hooks/useThresholdsInfrastructureData.ts, with frontend-modern/src/features/alerts/thresholds/hooks/useThresholdsData.ts limited to composing them. Thresholds-table controller state lives in frontend-modern/src/features/alerts/thresholds/hooks/useThresholdsTableState.ts, so future cleanup should extend those feature hooks or tab owners instead of rebuilding resource normalization, tab render surfaces, or thresholds-table runtime state inside the shell component.

The alerts page now also applies the same shell-versus-feature rule to configuration orchestration. frontend-modern/src/pages/Alerts.tsx is the page shell, while frontend-modern/src/features/alerts/AlertsConfigurationSurface.tsx is the feature shell. The canonical runtime owner is now frontend-modern/src/features/alerts/useAlertsConfigurationState.ts for alert config transport and org-switch reload orchestration, frontend-modern/src/features/alerts/useAlertsConfigurationSnapshotState.ts for the default-backed mutable configuration snapshot plus apply/capture/reset ownership, frontend-modern/src/features/alerts/alertsConfigurationModel.ts for config normalization, factory defaults, docker-gap validation, and payload serialization, frontend-modern/src/features/alerts/alertOverridesModel.ts for override normalization and resource-backed projection. That shared feature-model boundary must also canonicalize legacy shared-storage override keys and hashed storage resource ids onto the storage metrics target id before thresholds rows are derived, so migrated Ceph/shared-datastore overrides and Ceph pool overrides survive the feature-shell path instead of dropping out of the live editor, and frontend-modern/src/features/alerts/useAlertOverridesState.ts for reactive override state and thresholds-facing resource selectors, and frontend-modern/src/features/alerts/alertDestinationsModel.ts for destination config normalization and payload shaping, and frontend-modern/src/features/alerts/useAlertDestinationsState.ts for notification destination reload and persistence orchestration. Within that alerts configuration runtime, canonical container-runtime projection now belongs to alertOverridesModel.ts, useAlertOverridesState.ts, and useAlertsConfigurationState.ts. The thresholds Containers workspace must treat API-backed app-container parents such as TrueNAS as first-class Container Runtimes, while Docker-only controls in ThresholdsTableDockerTab.tsx remain gated to real docker-host resources instead of leaking onto platform-managed runtimes. frontend-modern/src/features/alerts/useAlertWebhookDestinationsState.ts now owns webhook runtime, and the email and webhook destination forms compose the shared TagInput for resource-tag routing. Their feature state owns tagFilter plus the all/any mode, an empty filter presents global delivery, and webhook list cards render the saved tags without creating a second tag-input primitive or destination- local normalization contract. frontend-modern/src/components/Alerts/ResourceTable.tsx now follows the same shell rule: the shell only picks desktop vs mobile render ownership and bulk-edit composition, while frontend-modern/src/components/Alerts/AlertResourceTableDesktop.tsx, frontend-modern/src/components/Alerts/AlertResourceTableMobile.tsx, and frontend-modern/src/components/Alerts/AlertResourceGroupHeader.tsx own the render-heavy table/card/group-header surfaces. Shared runtime state remains in frontend-modern/src/components/Alerts/useAlertResourceTableState.ts, shared row rendering remains in frontend-modern/src/components/Alerts/AlertResourceTableRow.tsx, and shared metric normalization remains in frontend-modern/src/components/Alerts/alertResourceTableModel.ts. Desktop rows and mobile resource cards also share one alert-delay handoff: their timer action emits the canonical resource id plus the first supported CPU, memory, or disk signal, while ThresholdsTab.tsx owns selection state and AlertIntentPolicyPanel.tsx owns expansion, scrolling, field inheritance, and API persistence. The action remains available for non-threshold-editable rows and falls back to state.offline when no supported metric exists. New platform threshold sections must forward this shared handoff rather than add a platform-local delay editor or a second intent-policy transport. frontend-modern/src/features/alerts/useAlertDestinationsTabState.ts now owns destination test actions, retry orchestration, and delivery-health loading while frontend-modern/src/features/alerts/tabs/DestinationsTab.tsx stays the render shell and should compose the dedicated email, Apprise, webhook, and load/error section owners plus frontend-modern/src/features/alerts/AlertDeliveryHealthCard.tsx instead of carrying those panels inline. That card is a route-owned danger surface, not a global success badge: degraded retained terminal delivery state and unavailable health reads remain visible, while healthy and retry-pending state add no banner. Its operator copy stays centralized in frontend-modern/src/utils/alertDestinationsPresentation.ts. Future cleanup should extend the transport hook, config model, override hook, or destinations runtime hook based on the true owner, not move config control flow back into the top-level page shell. The alert email provider picker now also follows the shell/runtime split: frontend-modern/src/components/Alerts/useEmailProviderSelectState.ts owns provider-catalog loading and provider-default application, while frontend-modern/src/components/Alerts/EmailProviderSelect.tsx stays the render shell and should not re-accumulate NotificationsAPI.getEmailProviders or a second local email-config contract inline. The alert scheduling surface now follows the same shell-versus-section split: frontend-modern/src/features/alerts/tabs/ScheduleTab.tsx should compose the dedicated quiet-hours, cooldown, grouping, recovery, escalation, and summary section owners while frontend-modern/src/features/alerts/useAlertScheduleState.ts remains the canonical runtime owner. The same rule now also covers cross-tab incident timelines: the shared runtime owner is frontend-modern/src/features/alerts/useAlertIncidentTimelineState.ts, while frontend-modern/src/features/alerts/OverviewTab.tsx and frontend-modern/src/features/alerts/tabs/HistoryTab.tsx stay focused on surface composition. Future incident timeline fetch, note-save, or expansion control flow should extend that feature hook rather than forking back into either tab surface. Every surface invoking that owner must pass the canonical alert identifier and the occurrence start time as distinct arguments; a row key is local UI state, not a substitute for either API identity field. The shared IncidentTimelinePanel consumes loading, error, timeline, draft, and save state as accessors so an asynchronous result remains reactive across Overview, desktop History, and mobile History instead of freezing the values present when the panel first expands. Overview alert runtime now follows that same shell-versus-runtime split. The shell stays in frontend-modern/src/features/alerts/OverviewTab.tsx, while frontend-modern/src/features/alerts/useAlertOverviewState.ts owns derived alert stats, filtered ordering, and single/bulk acknowledge runtime behavior. Future overview control flow should extend that hook rather than restoring action timers or acknowledge mutations to the tab shell. Card, group and global acknowledgement requests use the owner passed from the shared realtime store through the Alerts page and overview hook. Busy controls therefore survive an overview remount. Partial group results expose both the success and failure counts, and each in-flight alert stays with its original request. The scope-lifetime and rendered-control proofs are frontend-modern/src/stores/__tests__/websocket.alertAcknowledgementRequests.test.ts and frontend-modern/src/features/alerts/__tests__/OverviewTab.acknowledgementbusy.test.tsx. The overview follows the shared-clock rule for relative ages: the alert card reads useRelativeTimeNow for its started age and the live reading's stale cut-off, and the hook reads it for the Last 24 Hours count. A hook-local minute signal used as "now" is a table-local interval under that rule; the hook keeps a minute interval only to refresh delivery diagnoses, a server read whose cadence is not tied to the age clock. Render-heavy overview ownership now lives in frontend-modern/src/features/alerts/AlertOverviewStatsCards.tsx, frontend-modern/src/features/alerts/AlertOverviewActiveAlertsSection.tsx, and frontend-modern/src/features/alerts/AlertOverviewAlertCard.tsx, so future card-list or timeline-card presentation work should extend those surfaces rather than expanding frontend-modern/src/features/alerts/OverviewTab.tsx back into a mixed shell. Alert history runtime now follows that same pattern. The shell stays in frontend-modern/src/features/alerts/tabs/HistoryTab.tsx, while frontend-modern/src/features/alerts/useAlertHistoryState.ts owns history fetch, persistent filters, history-clear behavior, and composition of the derived history owners. Resource-incident panel runtime now lives in frontend-modern/src/features/alerts/useAlertResourceIncidentsState.ts, while frontend-modern/src/features/alerts/alertHistoryModel.ts owns grouped/trend derivation and the bucket/range analytics contract. The render-heavy surfaces now route through frontend-modern/src/features/alerts/AlertHistoryFrequencyCard.tsx, frontend-modern/src/features/alerts/AlertHistoryFiltersCard.tsx, frontend-modern/src/features/alerts/AlertResourceIncidentsPanel.tsx, frontend-modern/src/features/alerts/AlertHistoryTableSection.tsx, frontend-modern/src/features/alerts/AlertHistoryTableGroupRow.tsx, frontend-modern/src/features/alerts/AlertHistoryTableAlertRow.tsx, and frontend-modern/src/features/alerts/AlertHistoryAdministrationCard.tsx. Future alert-history control flow should extend the hook, pure history analytics should extend the model, and section rendering should extend those owners rather than rebuilding any of those concerns in the tab shell. The resource-resolution handoff into the resource-incident panel now belongs to the history state rather than the tab shell. The history model also owns the close outcome: an alert that closed with a typed resolution is status moved (badge moved to agent, its account in HistoryItem.closeDetail) rather than resolved or acknowledged. The table row carries the account in the badge title and the phone card renders it under the message, while getAlertHistoryStatusPresentation and getAlertResolutionDetail in utils/alertIncidentPresentation.ts own the wording. Resource-change readers follow the same split: getResourceChangePresentation and formatResourceChangeHeadline in utils/resourceChangePresentation.ts own the Alert moved label, its neutral-blue tone and the summary headline for an alert_resolved change with alert_resolution metadata, and frontend-modern/src/components/Infrastructure/ResourceChangeSummary.tsx consumes that headline instead of composing its own kind prefix. frontend-modern/src/features/alerts/useAlertHistoryState.ts re-exposes the getResource resolver it is already given, and frontend-modern/src/features/alerts/AlertResourceIncidentsPanel.tsx reads it from there. The panel mounts inside the history row that opened it rather than beside the tab's other cards, so a prop chain from frontend-modern/src/features/alerts/tabs/HistoryTab.tsx would have to thread through the table section, the group row, the alert row, and the mobile list to reach it. The tab shell itself should still only react to the current alertData() contract rather than reviving deleted history-state aliases such as filteredAlerts(). The panel may render compact route chips, but it must stay on shared route helpers and feature-owned composition instead of growing provider-local routing logic or another page-local resource lookup path. Top-level settings surfaces must route through Settings.tsx, SettingsPageShell.tsx, and frontend-modern/src/components/shared/SettingsPanel.tsx instead of reintroducing bespoke outer page headers or one-off top-level panel framing. The shell metadata driving those surfaces is part of the same boundary as well: frontend-modern/src/components/Settings/settingsHeaderMeta.ts and representative top-level panels such as frontend-modern/src/components/Settings/APIAccessPanel.tsx, frontend-modern/src/components/Settings/AISettings.tsx, frontend-modern/src/components/Settings/AIModelSelectionSection.tsx, frontend-modern/src/components/Settings/AIRuntimeControlsSection.tsx, frontend-modern/src/components/Settings/AISettingsStatusAndActions.tsx, frontend-modern/src/components/Settings/AIProviderConfigurationSection.tsx, frontend-modern/src/components/Settings/AISettingsDialogs.tsx, and frontend-modern/src/components/Settings/aiSettingsModel.ts now also define the canonical AI settings runtime boundary. AISettings.tsx is the shell, frontend-modern/src/components/Settings/useAISettingsState.ts owns the runtime lifecycle and persistence flow, model/provider setup now routes through AIModelSelectionSection.tsx, budget and timeout controls route through AIRuntimeControlsSection.tsx, and readiness plus save/test actions route through AISettingsStatusAndActions.tsx. AISettingsStatusAndActions.tsx may expose provider connection status and test actions only for the Provider & Models page; section pages reuse the save bar without making Patrol or External agents look like provider setup screens. Future AI settings work must extend those section owners instead of re-inlining large runtime subsections into the shell. Provider-specific settings fields inside AIProviderConfigurationSection.tsx must remain model-driven through aiSettingsModel.ts extraFields, including Ollama keep_alive, so the shared provider panel owns framing, labels, help affordances, helper copy, and form binding instead of adding provider-local bespoke controls. That same AI settings boundary now also owns frontend-modern/src/utils/aiSettingsPresentation.ts, so shared loading, empty, OAuth, action/error, shell-description, and workload-discovery copy for the settings shell stays on one governed helper instead of drifting back into section-local strings. SecurityAuthPanel.tsx follows the same rule for post-auth restart guidance: it renders the getSecurityAuthRestartInstruction() label and optional command once for every deployment type instead of hand-written per-type blocks. frontend-modern/src/components/Settings/AuditLogPanel.tsx, frontend-modern/src/components/Settings/AuditWebhookPanel.tsx, frontend-modern/src/components/Settings/GeneralSettingsPanel.tsx, frontend-modern/src/components/Settings/NetworkSettingsPanel.tsx, frontend-modern/src/components/Settings/NetworkBoundarySettingsSection.tsx, frontend-modern/src/components/Settings/networkSettingsModel.ts, frontend-modern/src/components/Settings/SecurityAuthPanel.tsx, frontend-modern/src/components/Settings/SecurityOverviewPanel.tsx, frontend-modern/src/components/Settings/RecoverySettingsPanel.tsx, frontend-modern/src/components/Settings/SSOProvidersPanel.tsx, frontend-modern/src/components/Settings/useSSOProvidersState.ts, and frontend-modern/src/components/Settings/ssoProvidersModel.ts now also define the canonical SSO provider settings runtime boundary: SSOProvidersPanel.tsx is the shell, useSSOProvidersState.ts owns the reactive/API lifecycle, and ssoProvidersModel.ts owns provider-form normalization and payload building. That boundary must keep SAML creation on the same first-class action path as OIDC. SSOProvidersPanel.tsx may show read-only state from settings capabilities, but it must not render a self-hosted Pro upsell, UpgradeLink, or advanced_sso feature probe before opening the SAML provider modal. useSSOProvidersState.ts must treat provider type as form state only; SSO entitlement truth belongs to the backend/runtime capability contract, where OIDC, SAML, and multi-provider SSO are Community-tier capabilities. The group-to-role mapping form must retain IdP group names with embedded spaces through provider detail, edit, payload, and reload. Its entries are comma- or newline-delimited group=roleId pairs; whitespace within a group name is not an entry separator. Allowed Groups use the same comma/newline boundary so no-op saves preserve those exact names too. Keep the existing whitespace parsing for OIDC scopes, allowed domains and allowed emails separate from group-name parsing. ssoProvidersModel.test.ts and SSOProvidersPanel.test.tsx pin this round trip, and browser verification must inspect the saved mapping in the desktop and narrow edit dialog. frontend-modern/src/components/Settings/UpdatesSettingsPanel.tsx must keep page-shell titles, descriptions, and lead panel framing aligned instead of letting navigation/header labels drift away from the actual settings surface. The self-hosted Pulse Pro settings navigation item and route header metadata for frontend-modern/src/components/Settings/settingsNavCatalog.ts and frontend-modern/src/components/Settings/settingsHeaderMeta.ts are part of that same shell boundary as frontend-modern/src/components/Settings/ProLicensePanel.tsx and the shared settings billing presentation owner in frontend-modern/src/components/Settings/selfHostedBillingPresentation.ts; the system-billing navigation label, header title/description, and billing shell framing must all route through SELF_HOSTED_PRO_BILLING_PRESENTATION instead of drifting independently. The owned split is now explicit: the navigation label comes from navLabel, while the route header and billing shell reuse shellTitle plus shellDescription, so the settings IA and page shell stay aligned on Plans & Billing without reintroducing local label drift. That same settings-shell framing boundary also covers adjacent top-level settings references to the self-hosted commercial surface. When InfrastructureWorkspace.tsx or other settings-shell surfaces point operators toward Plans & Billing for billing, license status, Patrol mode, or paid feature access, they must reuse the shared referral copy from SELF_HOSTED_PRO_BILLING_PRESENTATION rather than drafting local “go there for billing” variants. That same shared presentation owner now also carries the entitlement-first commercial summary contract for self-hosted settings. The top-level navigation entry stays product-IA owned through navLabel (Plans & Billing), while the page header and shell title stay owned through shellTitle (Plans & Billing), and the billing shell must foreground the active plan name plus available capabilities before secondary billing or recovery detail. Paid upgrades should be able to confirm “Current plan: Pulse Pro” immediately after activation without hunting through generic billing language or a second page-local summary card model. That same shell boundary also has to stay safe for hosted tenant bundles. Settings-shell framing copy for self-hosted billing must route through selfHostedBillingPresentation.ts, with settingsNavCatalog.ts, settingsHeaderMeta.ts, and adjacent hosted settings shells consuming that settings-owned adapter instead of importing generic commercial presentation helpers in ways that can reintroduce top-level bundle-init cycles. frontend-modern/src/components/Settings/NetworkSettingsPanel.tsx is now a shell only for network-boundary controls. frontend-modern/src/components/Settings/NetworkBoundarySettingsSection.tsx owns the public URL, CORS, embedding, and webhook-boundary UI, while the editable discovery configuration entry point is owned by the infrastructure workspace instead of the System/Network route. Shared prop contracts for the network-boundary surface must extend frontend-modern/src/components/Settings/networkSettingsModel.ts instead of re-expanding the shell or reintroducing page-local section types. The Public URL control must describe both customer alert links and copied agent install/update command targets, recommend the externally reachable HTTPS domain used through a reverse proxy, and present request auto-detection as a fallback rather than implying the field affects notifications alone. The settings architecture proof pins that copy, and the responsive browser receipt proves the label and helper text remain readable without horizontal overflow. For hosted runtimes this is also a fail-closed artifact contract: the backend accepts only a canonically valid authoritative agent-connect URL or explicit Public URL and does not substitute auto-detection, direct Host, or forwarded headers. The same backend rule applies to the diagnostics Docker/Podman migration artifact before its token exists. Existing setup/install and diagnostics error presentation must therefore retain the backend failure and offer configuration recovery rather than synthesizing a copyable command or download from local form state. No frontend runtime file changes in this slice; registered-route proof lives in TestContract_HostedInstallerOriginsFailClosedAtRouter and TestContract_HostedDiagnosticsDockerPrepareTokenValidatesOriginBeforeMutation, while the existing settings architecture proof continues to own the Public URL guidance. The same backend boundary now covers returning-user and Stripe post-checkout magic links before their one-time token is persisted. The public request UI contract remains unchanged: unavailable hosted URL configuration produces the same generic accepted response for registered and unknown email and exposes no configuration diagnostic or account-existence signal, while checkout simply skips its optional sign-in delivery. No frontend runtime file changes in this slice; the API proofs are TestContract_HostedMagicLinkRequestValidatesOriginBeforeMutation and TestStripeWebhook_CheckoutMagicLinkValidatesOriginBeforeMutation. frontend-modern/src/utils/infrastructureSettingsPresentation.ts owns network discovery framing, scan-scope labels, subnet guidance and environment-lock copy consumed by DiscoverySettingsForm.tsx. Network discovery remains an Infrastructure feature, separate from the retired AI service scanner.

That same settings-shell boundary now also owns the shared settings presentation helpers that those panels consume. frontend-modern/src/utils/systemSettingsPresentation.ts is the canonical owner for shared system-settings presets, summaries, and customer-facing action copy, while frontend-modern/src/utils/ssoProviderPresentation.ts owns the shared SSO provider labels, empty states, and action/status messaging. Future settings copy changes in those areas should extend these helpers instead of inlining panel-local strings inside the shell or reactive state owners. Shared infrastructure action-link framing owns service-resource action links. frontend-modern/src/components/Infrastructure/serviceDetailLinks.ts returns only the PMG thresholds link today: the PBS recovery-events link it used to build was retired with the platform-first route migration, and PBS backups are reached through the Proxmox Backups tab. A future service recovery link must target the platform tab that owns those rows rather than a recovery-events workspace. Alerts, Patrol, and similar feature shells render no resource handoff chips since the cross-jump strips were retired on 2026-05-16 (see the retired-helper note above). A future cross-surface link must take its href and label from a route helper in frontend-modern/src/routing/resourceLinks.ts instead of freezing raw route strings, local link dedupe, or provider-specific link chips inside a feature panel; chip styling belongs in the feature shell. That same shared primitive boundary now also owns persisted column-identity migration for governed surfaces. When a v6 surface canonicalizes saved column IDs, frontend-modern/src/hooks/useColumnVisibility.ts must accept explicit legacy-to-canonical aliases so existing local preferences migrate forward without resetting user choices or forcing the runtime to keep deleted column IDs alive indefinitely. That same shared primitive boundary now also owns environment-lock presentation. frontend-modern/src/components/shared/EnvironmentLockBadge.tsx stays the reusable badge shell, frontend-modern/src/utils/environmentLockPresentation.ts owns the canonical badge label, title, and lock-button copy, and frontend-modern/src/components/Settings/DockerRuntimeSettingsCard.tsx stays the settings-shell consumer for environment-variable-locked container-update controls. Future environment-lock UX should extend those owners instead of reintroducing panel-local lock labels, badge styling, or title copy. The release-ready shell proof now also includes a representative desktop Playwright rehearsal in tests/integration/tests/15-settings-shell-consistency.spec.ts so general, organization, billing, relay, security, AI, updates, and recovery panels are all exercised through the built app shell under a seeded multi-tenant runtime. The security-facing settings panels within that shell now also follow an explicit shared boundary with security-privacy so shell framing stays here while auth posture, token controls, and privacy semantics remain governed as a trust surface instead of generic UX copy. That shared shell boundary now also covers version-matched docs-link framing: customer-facing privacy disclosures in shared settings surfaces must route through frontend-modern/src/utils/docsLinks.ts rather than panel-local external URLs. That same shared-shell framing also covers the concise telemetry summary in General settings. The shell may present the privacy contract in compact product copy, but the vocabulary for outbound usage telemetry must stay aligned with security-privacy: coarse deployment and lifecycle buckets, aggregate resource and outcome counts, coarse feature flags, and content-free Patrol, Assistant, and capability-API usage counters are allowed, while hostnames, credentials, infrastructure identifiers, URLs, paths, locale, browser events, prompts, chat messages, command text, action output, token values, and personal information are not. That same docs-link boundary also governs local legal docs surfaced from the settings shell: shared settings surfaces such as AIRuntimeControlsSection.tsx must route Terms-of-Service links through the shipped TERMS.md asset instead of hardcoding GitHub main URLs that can drift from the running build. The same shell boundary now also owns shared relay route framing copy: frontend-modern/src/utils/relayPresentation.ts is the canonical owner for the top-level relay settings description and availability copy used by both settingsHeaderMeta.ts and RelaySettingsPanel.tsx, so the route shell and its first SettingsPanel cannot drift into separate rollout or pairing descriptions or describe Relay as a Pro-only feature after Relay became its own self-hosted paid tier.

Single-surface settings pages that only render one canonical SettingsPanel must stay rooted directly at that panel instead of wrapping it in an extra page-level space-y-* container. frontend-modern/src/components/Settings/UpdatesSettingsPanel.tsx frontend-modern/src/components/Settings/RecoverySettingsPanel.tsx, and frontend-modern/src/components/Settings/AuditLogPanel.tsx are the current reference cases, and frontend-modern/src/components/Settings/__tests__/settingsArchitecture.test.ts locks that direct-root contract so single-surface pages do not quietly regain redundant outer spacing chrome. The same shared settings-shell boundary now also owns the API-backed source path inside Infrastructure. frontend-modern/src/components/Settings/InfrastructureWorkspace.tsx, frontend-modern/src/components/Settings/settingsHeaderMeta.ts, frontend-modern/src/components/Settings/settingsNavigationModel.ts, frontend-modern/src/utils/workloadEmptyStatePresentation.ts, adjacent setup guidance must use Add infrastructure as the operator-facing first-run label for API-backed onboarding, resolve that label to the shared Infrastructure destination and its inline ConnectionEditor add flow, and avoid reviving a standalone platform shell, Platform connections label, or provider-local route. That same settings-shell contract also owns the shared infrastructure summary state. frontend-modern/src/components/Settings/useInfrastructureSettingsState.ts, frontend-modern/src/components/Settings/useSettingsInfrastructurePanelProps.ts, frontend-modern/src/components/Settings/InfrastructureWorkspace.tsx, frontend-modern/src/components/Settings/useTrueNASSettingsPanelState.ts, and frontend-modern/src/components/Settings/useVMwareSettingsPanelState.ts must derive Proxmox/PBS/PMG/TrueNAS/VMware counts and availability from one shared infrastructure settings state source instead of letting the top-level ledger and inline credential flows fetch the same connection state separately. Phase 9 retired the standalone PlatformConnectionsWorkspace.tsx, TrueNASSettingsPanel.tsx, and VMwareSettingsPanel.tsx shells; they remain labels and proof history, not live presentation surfaces. That same shared settings-shell boundary also owns provider parity inside the inline add flow. Adding VMware may extend the same card, empty-state, dialog, and summary-shell patterns used by TrueNAS, but it must not introduce a VMware-only outer page shell, alternate settings route hierarchy, or another summary vocabulary for connection health and contribution counts. While VMware remains admitted rather than supported, shared settings primitives must render its source-picker card with the manifest-derived preview badge and keep supported-source empty-state copy from listing VMware as available now. That same shared filter-presentation boundary also owns infrastructure source continuity on active surfaces. Settings infrastructure and platform/runtime pages must keep known canonical source options such as truenas and availability visible when configuration or route context establishes them, even when current unified-resource results do not contain that source, so platform handoffs from settings and other surfaces do not flash back to generic host-only language while the operator is still in a provider- or endpoint-scoped investigation flow. That same shared feature-presentation boundary also owns storage disk-detail fallback messaging in frontend-modern/src/features/storageBackups/. Shared detail presenters must describe the actual capability or identity gap that prevents history from rendering, rather than reviving agent-install guidance on API-backed platforms like TrueNAS when the canonical disk metrics target is already the owning history path. That same shared chart primitive boundary now also owns physical-disk live I/O drawers. frontend-modern/src/components/Storage/DiskDetail.tsx must render read, write, and busy charts through HistoryChart plus useHistoryChartState, using the canonical physical-disk history resource id, instead of reviving diskMetricsHistory, a page-local ring buffer, or another storage-only live chart primitive for the same telemetry. The shared shell boundary now also includes frontend-modern/src/contexts/appRuntime.ts as the only neutral owner for app-level websocket and dark-mode consumption. Shared shells and primitives such as frontend-modern/src/components/Settings/Settings.tsx and frontend-modern/src/components/shared/TagBadges.tsx may consume that module, but they must not import @/App or recreate shell providers. frontend-modern/src/App.tsx owns provider placement; primitives own reusable consumption only. That same shared settings-shell and banner boundary now also owns demo-mode commercial suppression. frontend-modern/src/components/Settings/settingsNavCatalog.ts, frontend-modern/src/components/Settings/settingsNavVisibility.ts, frontend-modern/src/stores/sessionCapabilities.ts, frontend-modern/src/stores/demoMode.ts, frontend-modern/src/useAppRuntimeState.ts, frontend-modern/src/components/shared/HistoryChartOverlay.tsx, frontend-modern/src/features/patrol/PatrolIntelligenceBanners.tsx, and frontend-modern/src/features/patrol/PatrolIntelligenceHeader.tsx must consume one shared bootstrap truth from /api/security/status.sessionCapabilities.demoMode and hide billing tabs, trial nudges, monitored-system warning banners, dashboard upsells, Patrol upgrade CTAs, history-lock paywalls, and other public-demo commercial affordances when the browser is rendering a public demo runtime. Shared primitives must not perform their own ad hoc /api/health polling, response-header inference, hostname heuristics, or per-banner demo branching; the runtime bootstrap, shared session-capability store, and shared banner hooks stay on one canonical owner so suppression stays coherent across customer-facing surfaces. That same session-presentation boundary owns the non-promotional self-hosted v6 app posture. Settings navigation, shared upgrade links, trial banners, monitored-system warning banners, history-lock overlays, and paid-feature gate primitives must honor resolved presentationPolicy.hideUpgrade by hiding prompts by default on ordinary self-hosted installs. Direct activation/recovery routes may still render their owned content, but sidebar discovery, trial CTAs, plan-review links, plan upsells, and feature upgrade links must not appear unless an explicit handoff, hosted-mode policy, or active entitlement says they should. That same shared app-shell boundary now also owns AI bootstrap silence on non-AI routes. frontend-modern/src/useAppRuntimeState.ts, frontend-modern/src/App.tsx, and frontend-modern/src/stores/aiAvailability.ts must treat /api/security/status.sessionCapabilities.assistantEnabled as the only general-route AI availability fact, while non-AI settings panels stay off /api/settings/ai and /api/ai/* until an owned Patrol surface is actually open. Backend settings/model reads stay on frontend-modern/src/stores/aiRuntimeState.ts. The governed browser proof in tests/integration/tests/11-first-session.spec.ts must continue to assert that plain settings routes render without AI bootstrap traffic or console noise. Shared table, disclosure, and form primitives must also stay explicitly typed at the browser edge. Summary rows may memoize repeated pending-update reads, shared buttons must preserve discriminated disclosure props, toggle and a11y helpers must expose exact event signatures, shared rows must accept typed data-* props, and reporting-panel helpers must remain ES2020-safe instead of depending on feature-local casts or newer string helpers. Settings report scheduling follows the same shell/runtime/model split as the rest of the Reports panel. ReportingPanel.tsx owns layout and shared controls, useReportingPanelState.ts owns API lifecycle and save/run/delete control flow, and reportingSchedulesModel.ts owns schedule payload normalization, labels, default form state, and cadence formatting. Schedule scope selection must reuse the shared ResourcePicker and reporting catalog types rather than creating a separate resource selector or browser-local schedule API contract. The report list distinguishes saved files, SMTP acceptance, retry scheduling, terminal failure and unconfirmed delivery. SMTP acceptance must not be labelled as inbox receipt. Actionable delivery errors and the next automatic retry stay visible at narrow widths. Saved server paths and previous success times belong in the report details disclosure. A deliberate retry of unconfirmed delivery requires a duplicate warning. Active attempts and pending retries refresh from the server, and the browser never manufactures a delivery success. The existing advanced-reporting capability remains the access boundary. Proofs: frontend-modern/src/components/Settings/__tests__/ReportingPanel.test.tsx, frontend-modern/src/components/Settings/__tests__/reportingSchedulesModel.delivery.test.ts, frontend-modern/src/components/Settings/__tests__/useReportingPanelState.test.ts. The report list keeps status and Run/Retry visible at every width. Scope, recipients, file paths, Edit and Delete share the full-width details disclosure. Keyed store reconciliation preserves row identity, open details and keyboard focus across background refreshes. The schedule editor is loaded only when a schedule is created or edited, with a visible loading state. It shares the reporting field primitive with the main report builder. Edit-only code stays out of the initial ReportingPanel bundle, and the existing bundle budget remains enforced. The Reports browser proof covers the deferred editor and its full form after the chunk loads. Responsive reporting actions are verified through the rendered reporting tests and browser interactions. Architecture checks do not require the retired schedule table columns or separate mobile action menu. The settings navigation model now exposes a single infrastructure-systems sidebar entry for the infrastructure settings area. The former infrastructure-connections and infrastructure-install entries have been removed from SettingsTab, settingsNavCatalog.ts, settingsPanelRegistry.ts, and settingsNavigationModel.ts. No future additions to the settings nav may restore infrastructure-connections or infrastructure-install as independent tab identifiers; panel routing within the infrastructure area must use InfrastructurePanelStep in-page state instead of URL sub-routes. frontend-modern/src/components/Settings/settingsNavigationModel.ts owns the explicit routeability check that rejects retired infrastructure/workloads aliases before the settings shell mounts. useSettingsNavigation.ts may redirect /settings and still canonicalize current settings destinations, but it must not translate removed infrastructure subpaths into onboarding queries or derive Proxmox platform state from those paths. The shared frontend source/platform vocabulary now also includes availability as an agentless monitoring source and network-endpoint as the canonical resource projection. Source labels, badges, settings add-flow copy, and availability management copy must use shared presentation helpers instead of feature-local wording, so availability probes stay visually aligned with the Monitoring availability settings surface without pretending to be a host agent install or a platform API connection. Availability setup presets for pingable machines/devices, MQTT, ESPHome, or similar agentless endpoints must also stay on the shared settings form vocabulary: presets may fill target kind, protocol, port, and path defaults, but display badges and drawers still derive Availability and Network Endpoint labels from the shared resource presentation helpers rather than from preset-local copy. Infrastructure rows for those same agentless endpoints must surface probe evidence directly in the row, not just as a green status dot or an Availability badge. The shared row presentation must expose the probe method and latest latency or failure result once, inline in the agentless endpoint's metric slot, while keeping recent check timing and fuller failure context in the tooltip or drawer so operators can understand what was measured without duplicated row chrome. Known platform resources use that same compact presentation when availability is attached. AvailabilityProbeStatusCard is the shared detail primitive for Workloads and Docker host drawers; it renders the complete target, protocol, latest result, latency when relevant, evidence freshness, and last observation. Its fact rows, and the matching service/probe/target rows in AvailabilityProbeSuggestionCard, compose InfoCardKeyValueRow so phone layouts remain condensed while wide cards keep each value adjacent to its label. Plural attached checks render as repeated bounded cards from availabilityChecks, while the row keeps one compatibility summary. Expired successful evidence must render an amber Stale state with no green Responding normally copy, and a never-observed check must render Not checked. A matched machine or service carrying an attached projection must not appear as a primary row in the Machines Availability checks tab; the distinct source-owned network-endpoint for that configured check must appear there regardless of whether its correlation state is attached, standalone, ambiguous, or unresolved. The tab's summary card owns the one "Manage checks" link; the checks table toolbar keeps only "Add service/device check". Operational navigation for those agentless endpoints belongs to the frontend-primitives-owned Machines surface as a focused Availability checks tab rather than a new primary nav item. The page may show availability checks beside standalone Pulse Agent machines, but Settings remains the add/edit owner and the app shell must not add a separate top-level Availability destination. That Availability checks tab owns a URL-addressable view=fleet presentation alongside the existing table; table remains the default and q plus status filters are shared between both modes. Fleet tiles combine the canonical current-health projection with the bounded history batch: categorical state coverage is labelled in text as reachable, unreachable, indeterminate, or unknown, and latency is drawn only for reachable evidence so gaps cannot be misread as zero latency. A tile opens the existing ResourceDetailDrawer rather than a service-monitor-specific detail model. History failure must stay inside the fleet surface with explicit copy while current status and resource navigation remain usable. Desktop and phone layouts must keep every tile keyboard-operable, preserve the textual legend, and avoid horizontal clipping at fleet scale. In the table, a secondary MetadataBadge beside a cell's primary value shares one flex line with it: the value is shrink-0 and the badge min-w-0 with a truncating label, so the value never clips and rows keep the single height PlatformWindowedRows measures. A second line under the value would break that windowing. Passing flex to the badge alongside its own inline-flex left the display to stylesheet order, which clipped the availability result. The Machines page must not pretend its machine list is a generic overview: the default tab is Machines, the Machines table is only for Pulse Agent-backed resources with host telemetry, and the full availability-check row list belongs to the Availability checks tab. Its default disk column follows the platform host-table scan pattern: multi-disk machines render compact per-disk mini-bars so operators can quickly see disk count and pressure distribution, while sorting still uses the highest-usage operational filesystem, platform plumbing stays out of the visible disk set, and full per-filesystem labels remain in hover/detail affordances rather than turning the row into a raw mount browser. Servers, laptops, desktops, and comparable computers monitored only by agentless reachability checks may use targetKind=machine in the availability form, but they stay in Availability checks until a Pulse Agent registers and supplies CPU, memory, disk, and network telemetry. Machines empty and handoff actions must lead to Pulse Agent install or the Availability checks tab, not to an agentless machine row in Machines.

Mobile product layout is a shared primitive contract, not a page-local styling exception. At supported phone widths, tab rails and dense data surfaces must preserve their readable intrinsic width inside an owned horizontal scroll container; expanded inline detail must remain bounded by the visible viewport; and active destinations must be scrolled into view. Shared search, filter, disclosure, navigation, copy, and row-action controls must keep a 40-pixel mobile touch floor while retaining their compact desktop density. Settings navigation must own a viewport-bounded vertical scroll region so its route list does not push the active panel below the page. Settings data grids must also define a phone information hierarchy instead of depending on horizontal scrolling as their first responsive behavior. Identity, current state, and immediately available actions stay visible; lower-priority timestamps, external IDs, and verbose detail columns may be hidden through the shared symmetric column boundary. When a hidden field is still needed for the phone decision, its concise value belongs as mobile-only secondary context in the surviving identity cell. Multi-action cells may collapse visible labels on phones only when every action keeps an accessible name and the shared touch floor.

Patrol finding handoffs must derive approval posture from the canonical typed action state and approval policy, not only from a legacy approval id. A pending_approval action or any non-none approval floor remains explicitly approval-bound in shared handoff metadata so Assistant, the collapsed finding row, and the expanded action review cannot disagree.

The shared Actions dialog remains the responsive and accessible review primitive for typed APT maintenance. Its heading supplies the dialog accessible name, the close control has an explicit name, pending action rows are keyboard reachable, and the scroll-bounded panel keeps safety, execution, verification, recovery, delivery, and next-step content actionable at desktop and 390-pixel phone viewports. Responsive layout must not hide the exact parameter authority, evidence source, or recovery instruction, and it must not add a duplicate legacy action path or verification card when ActionResultV2 is present. Read-only sessions keep the review packet inspectable but must not render approve, reject, or run controls, while settled historical records must not be mislabeled as expired actionable reviews. For an aged, receipt-pending execution, the same responsive review dialog may expose an audit-recovery disclosure to an eligible local administrator. It must show the receipt-pending state to every viewer as an unknown outcome, with a non-mutating in-dialog re-read of the same action and a warning not to create a second plan while the first outcome is unknown. The re-read remains available before the recovery window and in read-only sessions; a failed read retains the unknown state and never implies that the action was sent again. Recovery must first instruct the operator to check the resource's actual state outside the action record, then require a written reason and an explicit acknowledgement before a fresh action read and guarded force-fail call. Desktop and phone layouts keep this confirmation visible without implying a retry, cancellation or failed mutation. Read-only and settled records show no recovery control; permission hints in the client never replace the server's authority check. Its action controls are also plan-identity-bound: a missing reviewed planHash renders explicit replan guidance and hides approve, reject, and run controls, while an actionable record sends the exact displayed hash on every mutation. The dialog is also route-backed through the canonical action query parameter. Contextual surfaces use the shared button-link primitive to hand off an exact typed action id; the Actions route opens that durable review directly, selects the matching Open or History subtab from server-authored lifecycle state, and removes the query when the dialog closes. Feature pages may summarize action context, but they must not recreate approve, reject, run, progress, or outcome controls outside the shared Actions review. The Actions route owns overlapping reads by request generation. A slower detail response or late dialog refresh must not replace a newer URL-selected action or reopen a closed review; a mismatched server action id is rejected. An older Open or History list response must not overwrite the newer tab's results or error. Actions.requestOwnership.test.tsx covers both response orders and close while a receipt re-read is pending; the browser navigation proof checks the rendered dialog and URL at desktop and phone widths without sending an action mutation. The Actions ledger is a peer top-level navigation destination. Patrol remains the primary detection and investigation home and may expose a route-backed Actions handoff, but Actions owns its pending-approval count and selected state. The canonical /actions route remains stable for exact action deep links and universal audit records originating from Patrol, Assistant, MCP, or manual controls. While that route is open, desktop and mobile navigation select Actions and the browser title identifies Actions. The first-class navigation entry composes the existing shared review dialog, route, API client, and durable action identity rather than creating another action client. When the trusted audit origin is present, both the queue row and decision packet show bounded product attribution such as From Patrol; unknown first-party surfaces fall back to From Pulse, and absent origin remains absent rather than guessed. The shared action review also exposes Open Patrol record only when a Patrol origin carries its canonical operationalRecordId; the link targets the existing route-backed Patrol attention selection and never derives identity from display copy, resource IDs, finding IDs, or action reasons. Older correlated Patrol actions may expose Open Patrol to the Patrol home, but never label that fallback as a record-specific return.

Protection posture presentation boundary

Platform coverage tables render the storage/recovery-owned four-state ProtectionPosture contract without inventing age, failure, verification, or coverage states from raw browser payloads. Protected uses the shared success tone, attention uses warning, unprotected uses danger, and unknown uses muted presentation. The compact row remains actionable: plain-language rationale and provider history/permission limitations live one disclosure deeper beside bounded restore evidence. A posture fetch failure must keep the evidence inspectable and show unknown, never a locally inferred healthy fallback. The table renders the server explanation verbatim and keeps its own fallbacks (checking, not rated, unknown) and the snapshot-only age tooltip in the same plain voice, without model terms such as canonical identity or provider evidence. The bounded restore evidence keeps the newest point and the newest completed point of each source before filling with the newest overall, so a run of guest snapshots or a running job cannot push the latest usable backup out of the disclosure. The coverage strip's posture rule sits behind a collapsed "What the counts mean" toggle under the legend, not an always-open paragraph, counter or badge row.

Protection table controls continue to compose shared filter, table, status, counter, loading, error, disclosure, and inline-detail primitives. The bounded batch hook retains fulfilled values during refresh and issues at most one request per 200 resource rows, rather than placing a request under each row. Table-local provider histories keep issue prominence inside that existing control rail through withPlatformAttentionCount: TrueNAS Protection carries the compact failed/warning count, vSphere Health carries the compact critical/warning count, and each built-in row order keeps attention ahead of routine activity. They must not add page-level attention summary cards above their toolbars. Full-width attention summaries are reserved for genuinely cross-section overview state, such as Kubernetes aggregating nodes, workloads, and health signals with distinct destinations. Running replication remains an ordinary event state rather than page-level alarm copy.

Operational Trust attention interaction boundary

The Patrol selected-detail surface owns the smallest lifecycle interaction set: acknowledge, return to open, temporary suppression with a required reason and one of the bounded 1-hour, 24-hour, or 7-day durations, and return to active. Every mutation refreshes both the shared detail and list projection. The queue retains screen-reader names, keyboard activation, focus restoration after node replacement, reduced-motion behavior, and a no-overflow phone layout. Raw evidence history, lifecycle timeline, provider limitations, and action audit remain one disclosure or shared review deeper; no platform table or Assistant surface recreates these controls.

Trust-gate state presentation

Shared frontend composition must preserve the typed domain state supplied by Patrol, alerts, storage, and unified resources rather than flattening it into a generic success/error or loading flag. After the backend accepts a manual Patrol run, the page leaves the transient Starting state using the accepted run identity and performs one bounded status/history reconciliation; a structured backend rejection remains distinct from a browser or network failure. Any action proposed from that run still hands off to the shared canonical Actions review instead of adding approve, execute, or retry-mutation controls to the Patrol feature. Patrol autonomy controls follow the same server-truth discipline: after any successful autonomy PUT, the feature reloads the canonical GET projection before rendering the selected mode. Compact paid-runtime acknowledgements are not frontend state and missing nested fields in them must not crash the page. An absent, malformed, or Go zero-time expiresAt value is rendered as no expiry; it must never become a year-one locale date. A real bounded future expiry remains visible beside the acknowledgement status.

Storage detail primitives render physical-disk collection truth explicitly: temporarily unavailable, provider/controller unsupported, and unexpectedly missing evidence use different copy, and an unavailable per-disk I/O stream must not render an apparently live chart or synthetic zero counters. Alert threshold primitives similarly carry the current canonical override ID through edit, save, reload, and refetch while legacy IDs remain read-only compatibility candidates. Domain ownership stays with Patrol intelligence, storage recovery, alerts, and unified resources; the primitive layer owns consistent rendering, accessibility, and handoff behavior only.

The same storage presenters own the last-known disk temperature treatment. A temperature whose collection state is not available renders through PHYSICAL_DISK_TEMPERATURE_LAST_KNOWN_CLASS (muted, dotted underline, help cursor) with its reason as the title and screen-reader "last known" text, never through the threshold colour classes, so a retained reading cannot look hot or healthy. Table, drawer and pool surfaces share that one decision in frontend-modern/src/features/storageBackups/diskPresentation.ts rather than each re-reading collection.temperature. That decision and its class now live in the small frontend-modern/src/features/storageBackups/diskTemperaturePresentation.ts module, which diskPresentation.ts re-exports, so the Machines table and the machine drawer's Thermals rows apply the same treatment without pulling the Storage presenter into their chunks. Those surfaces and the guest drawer's Physical Disks card import it directly. When the cell's value is a retained host-agent SMART temperature, it renders muted with a dotted underline and screen-reader "last known", as does a retained value in the guest card. Tooltip and Thermals rows say "(last known)", and the Thermals row carries the collection reason as its title. The shared drawer History reads the same decision. For a physical disk, resourceDetailDrawerMetricsHistoryModel.ts takes the reading from physicalDisk.temperature, beside the collection state that qualifies it, because websocket rows carry no top-level temperature for a disk. When that state is set and is not available (an empty or absent state predates the contract and stays current), getResourceMetricsHistoryCurrentMetrics offers no current smart_temp. A positive retained reading then reaches getResourceMetricsHistoryDeferredMetrics as a deferred metric: with no history points the legend says "last known", never "current", and the reason renders above the chart. TrueNAS disk rows follow it too: the drawer's Temperature row reads "(last known)" with the reason as title and no heat tone, and the one-line summary leaves a retained reading out because it has no room for the reason. The Proxmox outdated sensor setup notice reads the registry's per-node proxmox.sensorSetupOutdated verdict, which counts a SATA or SAS disk as having a temperature only when the reading was collected now (diskinventory.TemperatureCollected in internal/unifiedresources/proxmox_sensor_setup.go), so a retained one cannot hide the notice for a disk Pulse cannot read now.

The focused browser proofs are frontend-modern/src/features/patrol/__tests__/patrolRunAcceptance.test.ts, frontend-modern/src/components/Storage/__tests__/DiskDetail.test.tsx, frontend-modern/src/components/Storage/__tests__/useDiskDetailModel.test.ts, and frontend-modern/src/features/alerts/thresholds/hooks/__tests__/truenasThresholdPersistence.test.tsx.

Agent Doctor settings framing

Settings labels the application update panel Updates under the System group, whose description sends agent updates to Infrastructure, and keeps agent lifecycle triage in the separate Agent Doctor dialog. Platform update notices, Diagnostics, and Infrastructure rows use the canonical Agent Doctor route handoff instead of recreating installer or repair controls. The dialog may enrich the shared connections ledger with structured diagnostics, but it must preserve loading, unavailable, unsupported, waiting-for-auto-update, removed, warning, and critical states rather than flattening them into a generic update badge.

The canonical query is agentDoctor; agentUpdates remains a compatibility alias that opens the same dialog and does not create a second settings surface. Scoped connection IDs filter active rows without hiding removed-agent history from the unscoped view. Copy-command controls render only for a backend- and frontend-confirmed supported platform; unknown and unverified FreeBSD/pfSense states show bounded guidance with no executable command. The focused proofs are frontend-modern/src/components/Settings/__tests__/infrastructureAgentDoctorModel.test.ts, frontend-modern/src/components/Settings/__tests__/InfrastructureWorkspace.test.tsx, frontend-modern/src/components/Settings/__tests__/DiagnosticsResultsPanel.test.tsx, frontend-modern/src/components/Settings/__tests__/settingsHeaderMeta.branchcov0713.test.ts, and frontend-modern/src/utils/__tests__/updatesPresentation.test.ts.

Alert intent and three-state availability presentation

The Alerts thresholds surface composes the versioned intent-policy editor from the shared form, status, disclosure, and loading primitives. It must preserve field-wise inheritance: omitted fields inherit, while explicit false and zero values remain deliberate overrides. Save uses the displayed revision, reports revision conflict without replacing local edits, and refreshes from the server-owned document after success. Preview renders clear, expected-transient, pending-grace, and would-activate as distinct states and never presents preview as a write.

The powered-off default is presented and persisted as the guest resource-type rule so VM and LXC resources inherit it without changing node or agent connectivity. Blank means inherit, 0 explicitly means no wait, and the UI states both meanings next to the control. Duration fields accept only base-10 whole seconds from zero through 30 days; an enabled backup hard cap must be positive. Invalid, fractional, negative, or oversized values remain local, show an actionable error, and issue no API write. Disabling backup extension is an explicit enabled: false rule and remains separate from disabling a guest's powered-off alerts.

Availability controls expose UDP mode, request payload, and optional expected response only where valid for the selected protocol. Unified-resource presentation keeps indeterminate visibly distinct from reachable and unreachable: open-or-filtered UDP uses warning treatment and bounded evidence copy, never a green success tone or a fabricated latency. The primitive layer does not infer detector, operator-intent, or recovery truth.

The focused proofs are frontend-modern/src/features/alerts/__tests__/AlertIntentPolicyPanel.test.tsx, frontend-modern/src/features/alerts/__tests__/ThresholdsTab.test.tsx, tests/integration/tests/85-powered-off-tolerance.spec.ts, frontend-modern/src/components/Settings/ConnectionEditor/__tests__/AvailabilityTargetSlot.test.tsx, and frontend-modern/src/utils/__tests__/availabilityProbePresentation.test.ts.

Pool-health evidence presentation

TrueNAS storage detail composition presents canonical and native pool state, structured scrub or resilver progress, pool error totals, affected vdev role and topology, recommendation, evidence codes, and evidence source from the canonical resource payload. It does not derive failed-disk claims from a missing disk row or replace native path-only leaf identity with a guessed device name.

TrueNAS alert presentation deduplicates a native provider signal projected onto system, pool, and disk rows by provider, native ID, and code, preferring the most specific canonical resource. Distinct evidence codes remain distinct rows. All acknowledgement, suppression, history, and action handoffs continue through shared alert primitives; no provider-only alert shell or email action is introduced.

frontend-modern/src/components/Infrastructure/__tests__/resourceDetailDrawerTrueNASModel.test.ts and frontend-modern/src/features/truenas/__tests__/truenasPageModel.test.ts are the focused presentation proofs. The governed browser proof must open the shared resource detail surface and verify the same evidence labels in rendered UI.

SSO endpoint URLs are presented as copyable only when the server supplied one

frontend-modern/src/components/Settings/SSOProvidersPanel.tsx renders the OIDC Callback / Redirect URL and the SAML SP metadata and ACS URLs as CopyValueButton chips that admins are told to register with their Identity Provider. The backend may now legitimately omit those values when it cannot resolve a public URL for the deployment, so each block falls back to guidance text from getSSOEndpointUnavailableHint in frontend-modern/src/utils/ssoProviderPresentation.ts instead of rendering a copy affordance around an empty or wrong value. The panel must never embed a localhost endpoint URL of its own.

The OIDC edit modal distinguishes the two reasons the URL can be missing: for a provider being created it explains the URL is generated on save and copied from the provider card afterwards (the modal closes on save, so it cannot show the URL itself), and for an existing provider it shows the same public-URL guidance as the card. settingsArchitecture.test.ts pins the guidance owner, the absent localhost literal, and the corrected post-save copy; the rendered fallbacks and copy affordances are covered by frontend-modern/src/components/Settings/__tests__/SSOProvidersPanel.test.tsx.

Entitled application branding stays app-shell-owned

frontend-modern/src/stores/systemSettings.ts owns the narrow reactive runtime-brand payload loaded during authenticated bootstrap. The shared frontend-modern/src/AppLayout.tsx shell is the only owner of applying that payload to the centered header lockup and route-aware browser title. A custom bounded banner logo replaces the built-in mark; a non-empty display name replaces the Pulse wordmark, while a logo with an empty display name may stand alone. Kiosk mode keeps its existing hidden-header behavior.

The Appearance surface edits the already-canonical reportBranding object through BrandingSettingsCard; it accepts PNG, JPEG, or GIF files no larger than the persisted inline-logo boundary, previews the exact saved material, marks the shared settings form dirty, and provides an explicit remove action. It must not create a page-local branding cache or render configured values when white_label is unavailable. Focused proofs live in BrandingSettingsCard.test.tsx, AppLayout.test.tsx, and stores/__tests__/systemSettings.test.ts.

ZFS datasets extend the existing storage detail primitive

StoragePoolDetail remains the owner of the expandable ZFS pool presentation. When a pool carries optional datasets, its already-expanded detail region renders the canonical dataset name and formatted used, available, referenced, and mountpoint values. Empty dataset collections add no new panel, route, or navigation state. The presentation mapper owns byte formatting and missing mountpoint fallback so components do not reinterpret provider data.

Host GPU telemetry reuses shared metric and history primitives

The standalone machine table may expose typed host GPU utilization as a toggleable, sortable metric-bar column, but it must render through the shared MetricBar primitive and use the unified resource's existing metric key. The row model owns validation, maximum-per-host selection, and the per-device inspection title; the component must not invent vendor-specific probes or a parallel GPU table. The table and shared View column picker expose GPU only when at least one current machine has finite utilization evidence, so estates without GPU telemetry do not carry an empty default column or an inert toggle.

The shared resource drawer keeps GuestDrawerHistory as the sole history renderer. Agent-backed hosts and explicit docker-host metrics targets, including standalone Unraid machines, Proxmox nodes, and Docker/Podman hosts, consume the single frontend-modern/src/components/shared/hostMetricsHistoryModel.ts HOST_METRICS_HISTORY_GROUPS catalog so CPU temperature history cannot drift out of one host surface while remaining on another. Host resources provide the current canonical CPU temperature as the initial fallback; SMART disk temperature history remains on the physical-disk resource rather than being mislabelled as host CPU temperature. Resources with typed GPU sensors extend the applicable host groups with core utilization, VRAM pressure, and GPU temperature series and provide current typed readings as initial fallbacks. API-only Proxmox nodes select their canonical node metrics target and reuse a source-aware subset of that catalog: utilization, network, and thermals remain, while host disk throughput is omitted because the PVE node API does not expose that stream. The drawer must not leave an unsupported disk-I/O chart in a permanent Collecting history state. Non-host workloads retain the default workload history groups unchanged. Storage metrics targets select a dedicated capacity catalog backed by the canonical usage series, while physical-disk targets select only the canonical busy, read, write, and SMART-temperature series. Neither target may inherit guest CPU, memory, or network charts that its backend history model does not record. Rendered table proof belongs in frontend-modern/src/features/standalone/__tests__/AgentsMachinesTable.test.tsx; drawer grouping and fallback proof belongs in frontend-modern/src/components/Infrastructure/__tests__/resourceDetailDrawerMetricsHistoryModel.branchcov0712.test.ts.

The Proxmox nodes table reads node history only

frontend-modern/src/features/proxmox/ProxmoxNodesTable.tsx draws its Trends sparklines through useWorkloadTableMetricHistory with series: 'nodes', so it polls the infrastructure summary and no guest history. The embedded workloads table below it owns guest history and passes series: 'guests'. ProxmoxNodesTable.test.tsx pins the option; the per-page polling budget belongs to performance-and-scalability.

Proxmox Storage reuses the shared product-family source scope

frontend-modern/src/features/proxmox/ProxmoxPageSurface.tsx mounts the canonical shared Storage surface for its Storage tab and supplies the hidden proxmox-all filter. Shared storage source matching owns that internal umbrella and admits normalized PVE, PBS, and other Proxmox-family source keys; visible source-picker filters remain exact. Agent-only physical-disk telemetry whose canonical parent is a Proxmox node is admitted through its explicit platform membership rather than being hidden because its fact source remains agent. The page must not create a PBS-only disk table or duplicate storage state in the Proxmox feature.

The rendered route contract in frontend-modern/src/features/proxmox/__tests__/ProxmoxPageSurface.contract.test.tsx and the shared-surface source guard in frontend-modern/src/features/proxmox/__tests__/ProxmoxBackupsTable.test.tsx pin this composition boundary.

ZFS dataset table rows inside StoragePoolDetail inherit their separator from the shared STORAGE_DETAIL_ROW_CLASS presentation constant. The component must not reintroduce a raw border-token class for dataset rows.

Coverage strip exception groups are filter toggles

ProxmoxBackupsCoverageStrip renders a BackupVerdict as a TableCard in the platform overview's vocabulary: the TableCardHeader strip carries the tone dot and "0 / 929 protected" (the headline sentence stays as screen-reader text), with the caller's tail muted on the right from sm. Each exception group is one 32px row of dot, reason, count pill and a ProgressBar for its share of rated targets, and a green Protected row closes the list. With onSelectGroup, a group row is a button with aria-pressed whose accessible name is its count and reason ("163 no successful backup"); workload names are a hover title only, because the filtered Coverage table lists them. Without onSelectGroup a row stays plain. The card owns no filter state; its caller maps a group key to a filter.

The settings nav gates Infrastructure, and the blocked-route fallback is capability-aware

infrastructure-systems now declares requiredCapability: 'infrastructureRead' in frontend-modern/src/components/Settings/settingsNavCatalog.ts, so shouldHideSettingsNavItem and shouldBlockSettingsRouteItem withhold both the sidebar entry and the route from a session the backend says cannot read it.

This is a different gate from system-relay and support-reporting. Paid feature navigation is hidden from ordinary free sessions, but its panel-owned direct route remains available for an explicit activation or recovery handoff. Infrastructure has no such route exception: every endpoint behind it is RequireAdmin, so a non-admin got an all-empty page whose pollers logged a warn-level denial on every tick. Hiding and blocking the item is what stops those pollers mounting.

The same rule now covers the admin-only System tabs. system-network, system-updates, and system-recovery declare requiredCapability: 'systemSettingsRead' in the same catalog, so both the sidebar entry and the route are withheld from a session that cannot read settings. They share Infrastructure's rationale rather than the paid-feature one: a free install can act on a paid tab by upgrading, but a non-admin cannot grant themselves admin, so the tab can only end in a panel they will never populate.

system-general is deliberately excluded. Theme, language, and unit preferences on that tab are user-scoped, so gating it would take personal settings away from every non-admin, and the panel is not empty for them. Proof: frontend-modern/src/components/Settings/__tests__/systemNavCapabilityGate.test.ts, which pins the withheld, granted, and unresolved cases plus the system-general exclusion.

Because DEFAULT_SETTINGS_TAB is infrastructure-systems, the blocked-route fallback in useSettingsAccess.ts can no longer resolve to the constant — that sent a refused session straight back to the tab that had just refused it. The fallback uses an explicit preference order: the default when reachable, then the user-scoped system-general tab, and only then the first remaining route. Catalog order is not a safety policy; after the admin-only routes are removed it can otherwise land a viewer on Plans, an upgrade surface they cannot administer.

The capability rule also covers every remaining panel whose mount read is admin-only: Availability checks, all three Pulse Intelligence tabs, Diagnostics & Health, Data & Reports, and System Logs declare their own named capability. Data & Reports keeps its independent advanced_reporting feature gate because feature discovery and administrative reachability answer different questions.

Navigation remains stable while /api/security/status is loading, but canMountSettingsPanel refuses to instantiate a capability-gated panel until the exact named capability is true. A failed or incomplete status response is a resolved denial: the route is removed and the fallback selects General. This prevents first-paint panel fetches from racing the capability request while keeping a successful status usable during later refreshes.

Pinned by the infrastructure-systems block assertion in settingsArchitecture.test.ts and by __tests__/infrastructureNavCapabilityGate.test.ts, which also pins that neither gate fires before the security status resolves — hiding on an unresolved status would flash the default tab away from an admin on every load. __tests__/adminOnlySettingsNavGates.test.ts and __tests__/useSettingsAccess.test.tsx pin the complete named-capability, fail-closed mount, failed-status, deduplicated-load, and General-fallback rules.

Infrastructure system names are plain in-app links

InfrastructureSourceManager renders a connected system's name as a plain <a href> to its monitoring page in both the desktop table and the phone cards; Solid Router intercepts it as an in-app navigation, so the component needs no router context. The name keeps its existing tooltip, carries a resting underline, and leaves the Actions column (budgeted for Manage and the install button) unchanged.

Alert monitoring actions preserve domain ownership

The Alerts overview may offer a compact per-resource Monitoring menu, but the menu is an adapter over the canonical resource operator-state API. It must preserve unrelated state on every write, including one-shot/recurring maintenance and descendant scope, distinguish availability-only expected-offline from all-attention mute, and state that retirement changes Pulse monitoring rather than deleting provider inventory. Resource detail and alert surfaces use the same typed monitoring and lifecycle vocabulary and provider-ownership presentation helper. They must not create local alert mute, archive, or removal state. The menu remains keyboard reachable, uses ordinary shared surface and border tokens, and retains usable controls at phone width.

The mobile navigation bar publishes its own height

frontend-modern/src/components/shared/MobileNavBar.tsx is the only owner of the bottom navigation bar's height. That height is content-driven and already includes the safe-area inset through pb-safe, so no consumer may derive it. The bar measures itself after mount, publishes the result as the --pulse-mobile-nav-height custom property on the document element, keeps it current through a resize observer and a window resize listener, and removes it when the bar unmounts. Measurement happens in onMount rather than in the element ref, because a ref runs before the node is in the document, where the measured height is zero.

Every surface that must sit on top of the bar reads that property: the Assistant overlay panel and its backdrop, the compact post-update notice, and the global GitHub star banner. Consumers must not add env(safe-area-inset-bottom) on top of the published value, and must not reintroduce a literal bar height. The declared :root value is a pre-measurement fallback only and deliberately under-estimates: reserving more than the bar's real height leaves a band between an overlay's backdrop and the bar that is neither dimmed nor click-blocked, while reserving slightly less is covered by the opaque bar. The star banner follows the published height at every viewport with a small gap above the bar, without a breakpoint override.

Proofs live in frontend-modern/src/components/shared/__tests__/MobileNavBar.test.tsx, frontend-modern/src/components/__tests__/GitHubStarBanner.test.tsx, and frontend-modern/src/__tests__/App.architecture.test.ts, which fails if any runtime source reintroduces a literal bar height.

Alert delivery log presentation

The destinations-tab delivery log renders through the shared Card primitive in the feature-owned AlertDeliveryLogCard. Outcome badges use the plain-language labels from alertDestinationsPresentation rather than queue vocabulary ("Failed, retries exhausted", never "dead letter"), failure detail lines use the shared red emphasis tokens in both themes, and entry rows wrap without horizontal overflow at mobile widths. The unavailable state is a role="alert" message distinct from the empty state, because "cannot read the log" and "no attempts" mean opposite things to someone deciding whether to trust their alerting. The card renders immediately after the delivery-health warning and recovery controls, and the Overview warning uses the shared ButtonLink primitive to navigate to the Notifications route. Attempt and held-event rows use semantic time elements with visible absolute local timestamps for timeline correlation and retain relative time only as hover context. The explanatory copy names the seven-day completed and 30-day dead-letter windows separately.

Storage rows distinguish retained observations

Storage table state presentation consumes the storage record's canonical freshness field before provider-native health labels. When polling fails and the last-known capacity remains visible, the row uses the shared amber warning tokens, labels the observation Stale, and exposes the last successful refresh age in its title. Freshness presentation remains in the pure storage row model; the table component must not infer age from render time or restyle retained capacity independently.

Credential-bearing destination panels use replacement semantics

The feature-owned external-watchdog panel composes SettingsPanel without creating a second settings shell. A stored credential-bearing URL renders as an empty password input with a configured replacement placeholder and an explicit Remove action; it must not be inserted into the DOM, tooltip, status copy, or client logs. Only a newly entered value may be revealed with the panel-local Show/Hide control. Status badges use shared theme tokens, error and unavailable states remain textually distinct, and the four-part status grid collapses without horizontal overflow at phone widths. This pattern is the required primitive composition for future secret-bearing destination panels.

Alert destinations share one severity-policy primitive

Email, webhook, Apprise, and entitled mobile-push panels compose the feature-owned DestinationSeveritySelect rather than implementing separate labels, option vocabularies, or responsive layouts. The control builds on the shared FormSelect primitive, exposes one associated label, and uses the same All alerts / Critical alerts only presentation at desktop and phone widths. Destination-specific help may explain transport semantics—mobile copy states its privacy and current-state boundary—but it must not redefine the policy. Alert feature state owns persistence and entitlement gating; the primitive owns presentation only. alertDestinationsPresentation.test.ts pins the shared vocabulary and the distinct mobile guidance.

Switched-off destinations collapse to one line

AlertDestinationOffSummary is the shared off-state body for destination panels: one muted sentence saying what turning the destination on does and a "Show settings" link button that reveals the panel's form without changing its enabled state, then moves keyboard focus to the first revealed control because the button itself unmounts. Email and Apprise panels compose it instead of rendering a full greyed form, so the panel header and its enable toggle keep their place while the page stays short. Apprise keeps its header test action; email's test action returns with its form.

Escalation configuration uses destination identity, not channel aliases

The alerts-owned escalation section composes shared settings, toggle, and form controls while presenting a feature-owned checkbox catalog keyed by logical destination ID. It sources that catalog from the same loaded destination state used by the Notifications tab, preserves selected-but-disabled and deleted entries visibly, and prevents the final selection from being removed. The critical-repeat control exposes a bounded numeric interval only when enabled and states every lifecycle condition that stops paging. Desktop and mobile layouts must retain associated labels and avoid horizontal overflow. Escalation level delays and repeat intervals share the rendered 5–180 minute bounds; the feature state clamps typed values before they can leave the control surface.

Alert overview rows and paused-delivery copy localize through the catalogs

The alert overview summary (alerts.overview.summary.*), the row disclosure (alerts.overview.action.moreActions / lessActions) and every string of the shared paused-delivery banner (alerts.delivery.paused.*, used by both the overview and the destinations tab) route through alertOverviewPresentation or alertDestinationsPresentation and the English, German and Spanish catalogs. Relative alert ages format with Intl.RelativeTimeFormat in the active locale rather than the English-only formatRelativeTime. Proofs live in frontend-modern/src/i18n/__tests__/i18n.test.ts and frontend-modern/src/utils/__tests__/alertOverviewPresentation.test.ts.

Page titles name what the page holds

A page title or its localized catalog entry names the job the page does, not a neighbouring feature: the alerts Schedule page is "Delivery Schedule" in every locale, and i18n.test.ts refuses a maintenance word in any locale's title because maintenance windows live elsewhere.

Repeated parent cells show once per group

When a table keeps one row per child so each child keeps its own figures, toggle and link, the parent's own readings render once, on the group's first row, and later rows leave the parent cells empty, so one parent never reads as several identical ones. The rows must already be grouped. The Backups servers table no longer uses this rule: since 2 November 2026 each row is named by its datastore with the server beside it, and carries the server's status itself.

A narrow table folds a dropped column's problem into one it keeps

When a responsive table drops a column at phone width, a problem that column would have shown must not vanish into the row's drawer. It takes the closest kept cell that answers the same question, in that column's tone, and only while it outranks what the cell normally says. The Backups "Backup server" table drops Jobs below 520px, so a reachable server's failed or late datastore job ("Sync failed") replaces State on that row, while an unreachable server keeps its connection state. The wording stays short enough to fit whole ("late" rather than "overdue"), because a truncated warning is not read.

Fixed-layout tables give free width to the column that varies

In a fixed-layout table whose columns change with the container, columns whose content has a fixed size (a time, a badge, a row of buttons) take exact pixels, columns of truncating names take a plain percentage, and the column that says what happened takes no width so it absorbs the rest. Percentages for every track hand most of each extra pixel to content that already fits. Hidden tracks are collapsed and also zero-width, with the table at 100%: WebKit (Safari) keeps a collapsed column's width in a fixed layout, so widening the table to compensate scrolls sideways there, and display: none on a column makes Chromium lay it out as an auto column that shares the free width. A max() or calc() of a length and a percentage on a column is laid out as auto. The Alerts History table follows this; App.architecture.test.ts pins the pixel floors, zero-width hidden tracks and Message's 18% floor, and a WebKit and Chromium probe measured no horizontal overflow from 1280px to 1920px windows.

A refused setting reports the state the server kept

When a write is refused because the server owns the value (an environment lock), the surface reads the effective value before speaking: no message when the server already holds what the operator chose, a message naming the variable when it holds the opposite, and the generic failure only when the state cannot be read. The setup wizard's usage-statistics opt-out follows this, and catalog copy keeps the variable name untranslated (i18n.test.ts).

A switch keeps one name in every state

A toggle's visible label names the setting, not its current state, and that label is the control's accessible name. A label that flips with the switch ("Notifications paused" beside an off switch) reads as the opposite of the truth. State goes in the switch position and, where it needs explaining, in separate status copy such as a banner. The Alerts overview Send notifications switch follows this (alerts.activation.switchLabel), and i18n.test.ts rejects state words in that label in every locale.

Status copy claims only what the state proves

When one stored state covers several histories, its copy states only what that state proves now. The alerts paused banner says notification delivery "is turned off", not that it "has not been turned on for this install yet", because the same pending state follows a deliberate switch-off.

Alert overview rows place Investigate by the shared breakpoint

AlertOverviewAlertCard reads the shared useBreakpoint() hook to decide where the Investigate split button renders: on the row from sm up, inside the More disclosure below it. It renders one instance rather than two CSS-hidden copies, so no menu id is duplicated, and it adds no resize listener of its own. Snooze never moves, so its dialog state survives a breakpoint crossing. The More button keeps its existing catalog labels and sets aria-controls only while its region is rendered.

Alert cards deep-link guests through the workloads route search

AlertOverviewAlertCard builds a guest alert's link with buildWorkloadsRouteSearch({ resource }) and passes the alert's resource id unchanged, or for a per-filesystem alert the backend-stamped guestResourceId of the guest that owns the filesystem, also unchanged. It adds no identity mapping of its own: the workloads selection state resolves non-row ids through the resource registry, and the summary focus bridge reveals the selected row, so the alert link and every other ?resource= link share one resolution and reveal path.

Alert cards deep-link storage through the storage route search

AlertOverviewAlertCard builds a Proxmox storage alert's link with buildStorageRouteSearch({ resource }) on /proxmox/storage, preferring the alert's storageResourceId metadata, and adds no identity mapping of its own. ProxmoxPageSurface passes routeActive to the Storage surface it keeps mounted behind its other tabs, so a guest tab's ?resource= link is not read as a storage deep link.

Record-to-resource links resolve through one hook

A surface that links a record (a Patrol item, a governed action) to its resource's page resolves the id with useResolvedResource and builds the link with buildResourcePagePath, rather than issuing its own ResourceAPI.getResource call. The action review's Open resource and Open Patrol links share one wrapping row whose links keep 44px targets on phones. A dialog offers a link away only when its host keeps the dialog in the address, so Back returns to it (showResourceLink).

Thresholds search is a route search

The thresholds resource search writes ?q= through navigate(..., { replace: true }) instead of a component signal, so a deep link, a reload and the back button all keep the narrowed list. Pages with an unsaved-changes leave guard must compare the destination path, not every navigation, before asking, or a route-owned search prompts on every keystroke (shouldConfirmAlertsLeave).

Alert groups use the shared localization boundary

Alert group disclosure and group acknowledgement copy routes through alertOverviewPresentation and the shared English, German, and Spanish catalogs. The presentation distinguishes a backend-declared shared-system relationship (linked signal(s)) from multiple detectors on the exact same resource (related) while keeping expansion and acknowledgement controls in the existing alert surface. Components must not hardcode this vocabulary or render the backend correlation reason or key as operator copy.

The labels retain singular/plural behavior, use the existing responsive button primitives, and introduce no parallel group component. Localization and presentation proofs live in frontend-modern/src/i18n/__tests__/i18n.test.ts and frontend-modern/src/utils/__tests__/alertOverviewPresentation.test.ts.

Infrastructure synthesis reuses that alert group and disclosure interaction rather than introducing an incident dashboard beside Alerts. A backend-declared infrastructure-incident group renders a compact summary immediately above its primary alert, labels the failure layer and whether the backend established a supported cause or only a related observation set, and exposes the backend reason, affected count, observation times, and bounded evidence IDs inside a native details disclosure. The existing linked-signal control expands every supporting detector card, so the operator can compare timing and challenge the inference without losing alert-level acknowledgement, snooze, timeline, monitoring-policy, or Patrol actions.

The summary must not derive relationships, choose a root cause, hide contradictory observations, or render an observation set as causal. Failure class and synthesis labels route through the shared English, German, and Spanish catalogs. AlertIncidentSynthesisSummary.test.tsx, useAlertOverviewState.test.tsx, and i18n.test.ts pin inspectability, uncertainty wording, and additive group behavior at the frontend boundary.

Repeated row controls carry row-specific names

A control repeated on every row of a table or card list (Timeline, Resource, an icon-only action) gets an accessible name that says which row it acts on, starting with its visible label so speech input still matches it, and a toggle reports aria-expanded. Localized names go through catalog keys with placeholders for the row subject (alerts.assistant.iconLabel, alerts.assistant.iconLabelAt); every locale keeps those placeholders (i18n.test.ts).

Layouts that swap a table for cards follow the container

A surface that swaps a table for a card list decides from the width of the element that holds them (useObservedElementWidth), with the viewport breakpoint only as the fallback before the first measurement, because side rails and panels leave a table far less room than the screen width suggests. Alerts > History switches below an 800px card (ALERT_HISTORY_TABLE_MIN_WIDTH).

A deep link narrows, then decides once

A route parameter that points a list at one record (Alert History's ?alert=) narrows the list to that record instead of scrolling a windowed list to it, waits until every source that could hold the record has answered before widening or declaring it missing, decides once per record so later filters stay the operator's, and offers one control back to the full list.

Mobile destinations are links and update evidence stays explicit

Fixed mobile destinations and menu destinations that navigate are anchors with real href values. Plain primary activation remains in the SPA so route warming and per-platform route memory continue to apply; modified activation, middle-click, and context-menu actions remain native browser behavior. Platform and overflow controls that open menus remain buttons. Disabled platform anchors resolve to infrastructure setup, and menu focus, Escape return, active state, badges, and narrow-layout containment remain unchanged. MobileNavBar.test.tsx, mobileNavBarModel.branchcov0712.test.ts, and AppLayout.test.tsx pin those distinctions.

The Proxmox node drawer presents checked zero, current positive, stale, unavailable, and not-checked package evidence as a labelled detail row with bounded copy and checked time. The compact table badge remains reserved for a current positive observation. Provider errors are never rendered. Component evidence is covered by NodeDrawerOverview.updateEvidence.test.tsx, proxmoxUpdateEvidence.test.ts, and ProxmoxNodesTable.test.tsx.

Denied update-list access exposes visible guidance beside that status through the shared HelpIcon and help-content registry. It explains that inventory and metrics use separate permission checks, preserves the choice to monitor with read-only access, and describes Sys.Modify as broader access that also permits refreshing package indexes. A denied response does not establish the exact missing privilege or an empty update list. Status text wraps through the shared detail-row contract at phone widths, including retained counts and checked time. The operator-context section and its help control retain their identity while readings and relative timestamps update. The shared help dialog remains inside the viewport with scrollable content, keyboard dismissal and focus return. Current-source browser receipts under frontend-modern/browser-verification/ cover desktop, intermediate, phone and short landscape layouts. These checks establish presentation and interaction, not installed Proxmox permissions or released acceptance.

An unavailable update check whose reason is permission_denied reads neutral, so it remains in Details rather than Worth checking. Other unavailable reasons stay warnings. Node Details values wrap rather than truncate. This preserves full reasons, agent IDs and CPU models at desktop and phone widths.

Denied update-list access exposes visible guidance beside that status through the shared HelpIcon and help-content registry. It explains that inventory and metrics use separate permission checks, preserves the choice to monitor with read-only access, and describes Sys.Modify as broader access that also permits refreshing package indexes. A denied response does not establish the exact missing privilege or an empty update list. Status text wraps through the shared detail-row contract at phone widths, including retained counts and checked time. The Summary section owns its help control while readings and relative timestamps update. The shared help dialog remains inside the viewport with scrollable content, keyboard dismissal and focus return. Current-source browser receipts under frontend-modern/browser-verification/ cover desktop, intermediate, phone and short landscape layouts. These checks establish presentation and interaction, not installed Proxmox permissions or released acceptance.

Time axes label round times

A chart axis built from fixed-width buckets anchored at "now minus the range" labels its inner ticks at the next round local time at their true positions, not at the raw bucket edges, and any label that shows an hour shows its minutes, so 24-hour locales never print a bare hour after a date (nextAlertAxisBoundary in alertHistoryModel.ts).

Docker drawers expose reduced helper coverage without actions

The Docker host drawer consumes the canonical optional collectionMode field. For typed-helper-summary it adds one bounded warning to the shared attention section and omits the container update management card; it does not fabricate zero update state or offer an action that the reporting collector cannot execute. Unknown or absent values preserve the direct-runtime presentation. Component and browser proofs cover the warning, action omission, mode transition, and desktop/narrow containment.

Patrol model choice carries guidance and a cost preview

The Pulse Intelligence settings surface answers "which model should I pick and what will it cost" at the point of choice instead of after the budget trips. The shared AIModelPicker accepts per-model annotations (badge, note, tone) rendered beside the model name and under its description in both pinned sections and provider groups, with the badge and note folded into the accessible option name. The Patrol model field and the shared default field (when no Patrol override is set) pin guided models in a "Suggested for Patrol" section, repeat the selected model's marker under the closed picker, and render the server-computed Patrol cost preview: monthly estimate, per-run assumption with tokens explained once, 30-day spend against budget, and the schedule recommendation. The Patrol schedule select prices each preset from the same projection and the schedule card explains a schedule that Pulse slowed for a per-token model; a schedule the install already chose is never changed. Dollars and prices never derive from model names in the settings surface; useAISettingsState fetches /api/ai/patrol/cost-preview and /api/ai/patrol/model-guidance and aiPatrolCostPresentation.ts owns the copy. AIModelPicker.test.tsx, AISettings.test.tsx, and settingsArchitecture.test.ts pin those distinctions.

Patrol weekly digest card is a read-only summary

The Patrol Activity tab gains PatrolWeeklyDigestCard ("This week") above Verified outcomes. It renders the server-computed GET /api/ai/patrol/digest rollup as definition-list stat tiles built from the shared Button and ButtonLink primitives and the existing surface, border, and muted text tokens; it introduces no new shared primitive, theme token, or layout helper. The only navigation it offers is the existing /actions route, shown only when Patrol-origin fixes are waiting for approval. Loading, failed-load, no-runs, and truncated-history states carry distinct copy, and a failed load never renders zero counts as if the week were quiet. Browser proof covers the desktop and narrow Activity tab in frontend-modern/browser-verification.json.

Phone Docker update labels preserve page-owned scrolling

Below the 33.999rem table-container boundary, Docker update cells reduce inline padding and wrap existing badge/button content within the allocated cell. The rule is scoped to docker-container-update-cell; it must not restore a nested horizontal scrollport or change app-shell touch handling. Phone platform wrappers retain overflow: clip. Labels remain text, not icon-only substitutes.

App.architecture.test.ts protects the CSS scope and scroll ownership; 96-navigation-socket-recovery.spec.ts checks real 390px text ranges and the non-scrolling wrapper. Browser receipts do not establish physical Android touch behaviour or every asynchronous update state.

Issue explanation follows the customer-journey qualification boundary, which separates scripted interaction proof from provider reasoning, remediation and repeatable customer outcomes.

Alert health attention preserves asynchronous ownership

The existing delivery-health card and shared buttons consume only the latest started health read's state. Configuration Retry can overlap a disabled card refresh; disabling that button is not a concurrency guard. Older completions must neither clear the latest request's busy flag nor replace its attention or unavailable presentation. Existing danger tone, accessible alert role, labels, confirmation and wrapping controls remain unchanged; no new primitive is added. The focused hook/caller tests and scripts/check-delivery-health-ordering.mjs cover this dependency at desktop and narrow widths using scripted health and queue-action responses, without claiming backend notification delivery.

Alert status distinguishes dispatch from destination evidence

Shared presentation preserves the alerts dispatch-evidence contract, including its source controls and qualification limits.

Resource incident reads retain lifecycle ownership

The resource incident hook gives each started read a unique per-resource owner. Only that owner may publish history, set the failed-read state, report a failure or clear loading. Reset invalidates all pending owners before clearing state; disposal invalidates them and prevents new loads. Overlapping reads for different resources remain independent. Closing a row still permits its in-flight result to populate the existing cache; reopening cached history and explicit refresh are unchanged. Requests are not transport-cancelled. No API, retention or notification-delivery policy changes. A retry clears the current failed-read state while retaining cached history until the owning request succeeds, and a superseded success cannot clear a newer failure.

The hook's ten ordinary regression/control cases cover success, catch and finally writes, reset/reopen and disposal. The existing panel tests cover its presentation. scripts/check-incident-request-ownership.mjs exercises the real hook and panel in Chromium at desktop and phone widths with scripted responses and fixture reset, overlap and unmount controls. It is component lifecycle acceptance, not an installed full-page or notification-delivery receipt.

The incident-history continuation also projects retained desktop expansion into the mobile drawer at the shared CSS breakpoint and keeps expansion state on return to desktop. The mobile action label describes the visible drawer. Operator note text and attribution survive the Assistant handoff while raw command output remains excluded. Shared event cards override inherited table no-wrap styling, and notes preserve line breaks. Long notes must remain readable in both inline desktop history and the mobile drawer. Final source-bound browser/model qualification is recorded in the customer-journey document.

The shared type-to-search registry excludes inputs in inert modal backgrounds, including prepared shortcut targets. Escape belongs to the active dialog and must not clear a background history filter or invalidate its return-focus target. Ordinary search shortcuts resume when the background becomes interactive again. When Assistant is already open, a desktop-to-phone transition must retain that destination rather than reopening the underlying history drawer above it.

Compact Organization header wrapping

The shared header and its controls must wrap below the small-screen breakpoint instead of forcing the document wider when an entitled Organization selector joins the session controls. Keep the selector, kiosk/logout controls and connection indicator available; do not hide overflow to conceal an inaccessible action. Desktop grid placement remains unchanged.

Proof: AppLayout.test.tsx pins wrapping and retained logout semantics. The signed-offline 05-settings-mobile-audit.spec.ts exercises real Organization, Access and Sharing routes at 320px and 390px, with app-shell width and full-scroll assertions; header screenshots retain the compact layout. This is independent of private RBAC implementation and hosted probation acceptance.

Alert history clear/read ordering

A successful history clear invalidates reads started before its completion and settles their loading state, so delayed responses cannot repopulate deleted history rows. Failed clearing leaves the pending read valid; later range refreshes remain available. This is view-state ordering, not a change to retention, active alerts, acknowledgement or notification recovery semantics.

The deferred-response cases in useAlertHistoryState.test.tsx verify successful and failed clearing plus subsequent range refresh. The registered scripts/check-incident-request-ownership.mjs browser proof also exercises the real history hook and administration card: load a row, start a pending range read, confirm clear, then release the obsolete response at desktop and phone widths. Scripted API responses establish component behaviour, not installed backend deletion or destination delivery.

Platform overviews may lead with their own incident table

A platform Overview that leads with its provider's health signals reuses that platform's existing incident table rather than a new summary card: the vSphere Overview renders VsphereAlertsTable with showToolbar={false} above its hosts table, only when signals exist, matching the TrueNAS Overview's health-alert table. Filtering stays on the platform's Health tab.

Kubernetes nodes state readiness in words

KubernetesNodesTable renders the node state from mapKubernetesNodeStatus as a desktop Status column beside the name (a shared StatusIndicatorBadge for anything other than Ready, muted text for Ready) and sorts NotReady first. The status dot stays decorative; a phone-only screen-reader label beside the name announces the state where the column is demoted under the duplicate-state rule, so no width announces it twice. The overview reuses the same table, limited to nodes needing attention and with its own persisted sort slot, rather than a summary card.

Large platform notices keep the inventory in view

PlatformOutdatedAgentNotice is maintenance guidance, not an incident, so it renders as one sentence in the info tone: the count, what the update brings, a keyboard-operable "Which hosts?" toggle (any count above one) that reveals every affected name, and the upgrade link inline in the same line rather than on an action row. This keeps the stale-agent notice from pushing the platform inventory and Storage search below the fold at desktop and phone widths, and from competing with real warnings for attention, while preserving every affected name on demand. The component test pins collapsed, expanded and collapsed-again states; 1440px, 768px and 390px browser checks verify placement and overflow. Each platform page names what the update brings in the reader's words through missingLabel: the latest fixes and machine, node, system or VM details. The notice selects hosts by agent version alone, and command execution is a separate setting the update preserves, so the notice must not promise Patrol actions or command support. Internal phrasing such as "agent-contributed detail", "agent-managed platform detail" or "in-guest telemetry" does not belong in the notice. The shared InlineNotice action text uses opaque 800-level colors for its four tones. The current Tailwind configuration overrides several 900-level palette tokens with 25%-alpha colors for translucent backgrounds, so using those tokens for notice links made a working action look disabled. The browser proof also follows the outdated-agent action to Agent Doctor with all 49 host IDs. The broader palette override should be corrected in its own shared-design slice, with background users migrated to explicit alpha utilities so other 900-level text consumers can use normal opaque color semantics.

Actions empty state follows the AI capability

getActionsWatchOnlyEmptyState takes an explicit aiEnabled input, which pages/Actions.tsx fills from the assistantEnabled session capability. While AI is off it returns no guidance, so the empty Open inbox keeps its plain copy instead of saying Patrol runs in Watch only mode or pitching Pro Patrol modes for a feature that is not running (issue #905). With AI on, the Watch only, switch and upgrade branches are unchanged. actionsWatchOnlyEmptyState.test.ts pins both states and the page wiring.

Pulse Mobile settings section label

The system-relay settings section is labelled Pulse Mobile in the nav catalog, the header metadata and every locale catalog; the product name is not translated. Its route id, feature gate and read capability are unchanged. Relay never provided remote access to the web UI, so no settings chrome may label it Remote Access.

Drawer History is inspectable without pointer hover

GuestDrawerHistory exposes a labelled native range input for groups with multiple stored observation times. Native arrow keys, Home/End and touch input select real stored timestamps; the control's value text includes the full local date/time and separately formatted series values. A series without a sample at the selected time remains missing rather than borrowing a neighbouring or live reading. The SVG has a linked textual description, including lone observations and the absence of stored data; a lone observation does not fabricate a trend.

Selection follows its timestamp across same-source refreshes, not an ordinal index that shifts when samples arrive. An expired selection snaps to an actual remaining observation. Changing resource type, id or range clears pointer and keyboard selection even when matching cached data arrives immediately. Empty, failed initial, absent-target and licence-locked views expose no inspection control. Existing matching-point retention and scoped retry remain unchanged. Mounted inspection regressions cover these boundaries. The direct-renderer browser fixture verifies native keys, focus, pointer coexistence and touch at desktop/phone widths across Chromium, Firefox and WebKit; it is not installed PBS collection or a screen-reader announcement-quality claim.

Drawer History pointer values share one observed time

Pointer inspection snaps to the nearest actual stored timestamp across the group's series, with equidistant observations resolved to the earlier time. Every displayed value and marker must belong to that timestamp. Missing series remain unavailable at the inspected time, not a nearest neighbour, latest observation or current reading. A lone stored sample, including zero, can be inspected without fabricating a trend. The SVG's dated description follows the same active time as the visible legend. Leaving the plot restores the normal latest/current legend; focused native inspection still takes precedence.

Matching refreshes reconcile the pointer with the current set of actual observations; resource/range replacement still clears pointer state. Mounted GuestDrawerHistory.pointer.test.tsx regressions cover sparse and disjoint series, zero rates, single observations, ties, focus precedence and replacement. browser-tests/history-pointer.cjs verifies the production renderer and query in Chromium/Firefox desktop and Chromium/WebKit phone emulation in both themes, including failed refreshes and source replacement. Scripted APIs establish presentation accuracy, not real PBS collection or installed #1723 acceptance.

Drawer History panels share a dated time window

GuestDrawerHistory uses one common time interval for every configured metric group, including utilization, network, disk I/O and thermals. The fulfilled API response's valid start/end interval remains in view even if it contains only a few minutes of observations. A sample at a given timestamp occupies the same horizontal position in every panel; a group's sparse coverage must not stretch independently to fill the selected range. Visible date/time endpoints and full local timestamps in their accessible labels distinguish overnight and multi-day windows. Native inspection still selects actual observations.

Returned edge observations widen the common envelope instead of being clipped or discarded, including aggregated bucket timestamps. An invalid API interval falls back to the valid observed envelope across configured groups, not a fabricated range. Non-date timestamps and unconfigured metrics cannot poison that geometry. A valid empty window has labelled endpoints but no trend or inspection control. Failed matching refreshes retain the window with its data; target/range replacement clears both until matching data arrives.

GuestDrawerHistory.window.test.tsx pins geometry, labels, refresh/replacement, empty/invalid windows and edge observations. browser-tests/history-window.cjs uses the production PBS table, resource drawers, History query and CSS with synthetic APIs, checking three separately mapped drawers and range/refresh behaviour in desktop and phone-emulated engines. This is presentation proof, not installed PBS/VirtualBox collection or a complete #1723 acceptance result.

Shared canvas History responses belong to their selection

useHistoryChartState cancels and invalidates old requests when the resource, metric, range, sampling cap, supplied-data mode or access state changes. Late successes and failures cannot replace the current selection's samples, loading or error state, even if cancellation is ignored. Selection changes clear old readings and hover state while loading; current initial failures remain visible. Matching supplied-sample refreshes retain pointer inspection; changing the selection clears it. Matching background refresh failures retain already loaded samples. Polls never overlap and stop for supplied data (including empty arrays), unavailable targets, locked ranges and unmount.

useHistoryChartState.test.tsx checks delayed success/failure, each selection field, polling, supplied-data transitions, locked/empty targets and cleanup. browser-tests/history-selection.cjs exercises the production canvas chart and accessible description in desktop Chromium and phone WebKit, including a late old-target response and current-target loading/failure. Synthetic response proof is not native PBS collection or whole-report #1723 acceptance.

Shared History keyboard inspection

The existing Storage pool and disk canvas charts accept keyboard focus with a visible focus indicator. Focus selects the latest observed sample; Left/Right step through actual observations, Home/End select the endpoints, and Escape or blur clears inspection without trapping focus. Each keyboard-selected timestamp and formatted value is announced politely. Pointer inspection remains available without announcing every mouse move. Matching sample refreshes retain inspection; selection changes and empty data clear stale readings. No readings are invented for missing data. Tooltips reserve room for a wrapped local timestamp and reading. The mounted HistoryChart regression and history-keyboard browser fixture cover navigation, bounds, focus exit, refresh and replacement. Synthetic browser proof does not establish native collector or assistive-device acceptance.

Shared History window control access

Each visible shared chart window selector is a named group for its metric (or "History window" when no label is supplied). Exactly the current period carries aria-pressed=true; switching one chart must not change another chart's selected state. Period buttons retain native keyboard/touch behaviour and explicit type=button, so changing the window cannot submit an enclosing form. A hidden selector exposes neither a group nor controls.

The header, summary and existing period choices wrap within their chart on narrow screens instead of forcing horizontal overflow. This changes neither available periods, subscriptions, History ownership nor sample interpretation. Mounted HistoryChart.test.tsx and the production-component browser-tests/history-window-controls.cjs cover selection, scope, native input, form non-submission and phone layout. Synthetic observations are not native PBS/NAS, physical device or assistive-device acceptance.

Shared History request-state honesty

An empty successful response states only that the selected time range has no samples; it does not claim collection is running or promise future readings. A failed matching background refresh keeps the last successful data and source, with a visible polite status also included in the chart's accessible description. The warning persists while retrying and clears only on success or selection replacement. This applies to previously empty results too. Initial failures remain separate errors. No polling frequency, request or access policy changes. useHistoryChartState.test.tsx covers failure, pending recovery, empty results and selection changes; browser-tests/history-status.cjs checks actual chart states in desktop Chromium and phone WebKit with synthetic transport failures. These checks do not establish native collection or released availability. HistoryChart takes enabled (default true) for a chart mounted inside a hidden drawer tab: false aborts its read and stops polling but keeps the readings it shows, and resuming the same selection reads at once without a loading state before polling again. A selection changed while paused still starts from nothing. The hook and mounted chart tests pin both paths. The point cap follows the chart's width but is not part of what the chart shows: a resize (or a full-page capture) re-reads the same series at the new cap while keeping its readings on screen with no loading state, and only pointer inspection resets, since its pixel positions moved. A changed target, metric, range, organisation or lock still starts from nothing. History axis labels must not crowd. The canvas chart draws as many evenly spaced time labels as its plot fits at the measured label width, from four down to the first and last time (getHistoryChartTimeTickCount), so a narrow live I/O card no longer runs "12:04 PM12:07 PM..." together. Both the canvas chart and the drawer charts blank a middle gridline label that repeats a neighbour (withoutRepeatedHistoryAxisLabels): a quiet link reads "<1K/s" once, not on two gridlines, and the gridline itself stays. The end labels always remain. guest-storage-2263.cjs records each chart's drawn time labels and checks they stand apart; pbs-history-refresh.cjs checks the drawer's network axis. The canvas HistoryChart draws the same min-max band the drawer charts do: when any bucket's lowest and highest readings differ (hasHistoryChartSpread), a 20% band spans them behind the line, so a short spike survives downsampling. It no longer fills the area under the line, so a storage chart reads like a drawer chart; a flat series is a line alone. A time label that repeats the one before it (a short window at a coarse label) is dropped (withoutRepeatedHistoryTimeTicks). Phone chart folds share one button, PhoneChartsToggle ("Show " / "Show fewer charts", aria-controls naming the folded group, hidden from sm up); the drawer History and the disk detail both use it. History names where its readings came from, from the reply's source, in one shared vocabulary (getHistoryReadoutLabel, getHistorySourceNotice): stored history (and demo data) reads "Latest stored ·

Shared History single observations

A single stored observation renders as a visible centred marker with one real timestamp label, including measured zero. It must not imply a trend or invent a duration. Pointer and keyboard inspection resolve to that actual observation; ordinary multi-observation interpolation and empty-result states are unchanged. The geometry's inverse timestamp mapping is shared by pointer inspection.

HistoryChart.sparse.test.tsx and historyChartModel.branchcov2.test.ts cover marker geometry, one timestamp label, pointer bounds, keyboard inspection and single/zero/multiple/empty refreshes. browser-tests/history-sparse.cjs checks actual series pixels and tooltips in desktop Chromium and phone WebKit, using the production chart and API transport with synthetic stored samples. This does not establish native collection, whole-report resolution or availability in a release. Existing request, source, units and access boundaries are unchanged.

Shared History touch inspection preserves native scrolling

A primary touch tap inside a Storage History plot selects the nearest actual observation at that horizontal position. Compatibility mouse focus must not replace that reading with the latest sample or announce it as keyboard input. The existing hover group and timestamp reconciliation remain shared with mouse and keyboard inspection. Single measured-zero observations remain inspectable; empty data and a changed selection cannot inherit an old touch.

Movement beyond ten CSS pixels, cancellation, multiple contacts and release outside the plot are not taps. Inspection neither prevents native touch defaults nor captures the pointer or changes touch-action: page scrolling and pinch zoom remain browser-owned. Blur and Escape clear inspection, and supported keyboard keys resume the existing dated, politely announced sample navigation.

Mounted HistoryChart.test.tsx regressions cover focus ordering, gesture rejection, selection ownership, refresh, empty/zero and input coexistence. browser-tests/history-touch.cjs exercises actual touchscreen taps through the production disk/pool details and API transport in phone Chromium/WebKit, native vertical panning in Chromium, and desktop mouse/keyboard controls. Synthetic sample presentation is not native collector, physical device, assistive-device or installed whole-report acceptance.

Shared Storage presenters preserve disk ownership

Pool-detail presenters under features/storageBackups/ must not use repeated device paths, UnRAID group labels or host display names as fleet-wide identity. Direct pool children are retained; inferred membership requires the same canonical host parent, or a complete native Proxmox instance/node pair when parents are missing. Known parent conflicts win over native names. Full ZFS device identifiers may normalize an optional /dev/ prefix but must not match arbitrary suffixes. Missing ownership produces no invented linked-disk facts. The storage adapter preserves the instance needed for that presentation scope; this does not rewrite canonical resources or History targets. Existing keyed pool details must update linked disks and local health through target/snapshot replacement without retaining peer-host rows. The mounted StoragePoolDetail.ownership.test.tsx and production-component browser fixture browser-tests/pool-ownership.cjs verify desktop/phone presentation and History coexistence; they are not field collection or native recovery evidence.

Scheduled report kinds

The Settings > Reporting schedule form composes the shared FormSelect for a "Report type" choice between the performance report and the Patrol weekly summary (patrol_digest). The choice renders only while the session's AI capability reports Patrol can run, or while an existing summary schedule is being edited, so an AI-off install sees the unchanged performance-report form. The summary variant hides the cadence, format, resource picker, tag filter, delivery, retention, attach, and save-copy controls and keeps name, weekday, time, timezone, recipients, and enabled; reportingSchedulesModel.ts owns the kind normalisation, the kind switch, the email-only payload, and the plain "Patrol activity, last 7 days" scope label. ReportingPanel.test.tsx, useReportingPanelState.test.ts, and reportingSchedulesModel.branchcov2.test.ts cover the gating, the reduced form, the pinned fields, and the saved payload. This is a presentation and form-state boundary; the schedule API contract, report rendering, and email delivery are unchanged.

Notifications activity has independent evidence-read states

The Notifications tab composes the delivery-log card with both attempt-read and held/deferred-event loading and unavailable state from its owned reader. Pending or failed evidence is never rendered as a confirmed empty activity window. Each failed source withdraws its own rows without hiding independently readable rows from the other source. Source-specific warnings have role="alert" and loading text has role="status"; a slow held-event read does not block the existing attempt-refresh button. Recovery clears a warning only after a successful read, without remounting the destination editors or changing their unfinished input.

The connected Notifications browser fixture uses the real tab/state/API/card chain at desktop and phone widths with keyboard/touch refresh, class-dark CSS, 403/503 controls and pending recovery. Unit controls separately cover current request ownership and disposal. No mutation, route, navigation entry, queue policy, permission or recipient-delivery contract is added by these states.

Drawer History does not interpolate across missing observations

The shared GuestDrawerHistory renderer connects a series only across consecutive stored times in that configured panel. If another series has a stored observation at an intervening time and this series does not, its line stops. Each isolated observation, including a measured zero, remains visible as a point; separate continuous segments retain all their original values, timestamps and common-window coordinates. Independently sampled series must not imply simultaneous observations or a continuous trend through absent data.

A visible, SVG-described explanation names the affected configured series and states the limit: missing readings do not identify an outage's cause or duration. There is no elapsed-time/cadence heuristic, synthetic zero, invented observation, new API field or collector change. Times outside the configured panel cannot split its series. Entirely absent series retain the existing missing/fallback policy. Live or retained snapshot carriers cannot repair a stored gap; only a matching stored response can. Source/range/access isolation, transient retained reads and exact-time keyboard/pointer/touch inspection remain in force.

GuestDrawerHistory.gaps.test.tsx and guestHistorySegments.test.ts cover split lines, singleton recovery, zero rates, interleaved sampling, multiple segments, configured-group isolation, invalid points, elapsed-time controls, live updates, stored refresh and denied/transient reads. GuestDrawer.test.tsx checks the full drawer's live-state recovery and unchanged request budget. Updated pointer/sparse regressions preserve their original missing-value and fallback obligations. browser-tests/history-observation-gaps.cjs exercises the production full guest drawer with synthetic APIs in Chromium/WebKit desktop and 390/320-pixel light and dark views. This is presentation acceptance, not native backup/QGA safety, provider outage attribution, installed recovery or release availability.

Threshold group headers carry a defaults line

CollapsibleSection takes an optional defaultsSummary rendered under the title as a muted line (data-thresholds-defaults-summary) that wraps only between items on phones and truncates from sm up, and is hidden while isGloballyDisabled is set, where it would promise alerts that cannot fire. Callers pass the shared summary helper's output rather than composing per-section strings.

Passive alternative for unavailable Proxmox guest disks (6 October 2026)

The existing disk guide offers already-present Machines filesystem readings and History as an interim view, with same-guest identity and metric-age checks. Recent agent contact and retained History do not establish current disk usage; working Machines readings do not repair Proxmox collection or prove correlation or thaw. No new agent installation, guest restart/restore, pause bypass or active probe is prescribed. Independent thaw, every-covered-filesystem writes and liveness remain the backup restoration boundary. The shipped mirror and backupSafetyHelp.test.ts validate the rendered guidance and anchor; both new controls reject the original parent guide. No runtime or wire contract changes.

PBS protection disagreement help (6 October 2026)

The existing PBS guide separates visible/matched/verified backup artifacts from subject-linked protection assessment and independent restore or guest recovery proof. Its passive Coverage check uses the shipped Job, History and Access labels, existing PBS records and scoped redacted reporting. No task is induced, no identity/permission/retention setting is changed and no policy cause is assumed. Frozen/unresponsive guests retain independent post-backup thaw, writes on every covered filesystem, liveness and only-prior-active restoration with the monitoring/alert outage disclosed.

pbsProtectionHelp.test.ts checks the production Markdown renderer, ordered passive steps, evidence distinctions, safety link/target, disclosure and exact shipped mirror. Exact-parent controls reject the old guide. This is safe help maintenance, not a collector/classifier repair, browser acceptance, native backup/recovery proof or published availability.

TrueNAS heat uses the configured disk policy (8 October 2026)

The production TrueNAS storage table and inline resource drawer pass each disk's canonical resource ID through the alerts store. Type defaults, a TrueNAS-wide value and the disk override produce one consistent status/reason and temperature tone; retained values stay muted and unjudged. The existing keyboard disclosure and phone table layouts are preserved.

Verification: TrueNASStorageTopologyTable.test.tsx, the drawer model tests, and browser-tests/truenas-disk-thresholds.cjs exercise desktop Chromium and phone WebKit, config refresh, Attention filtering and keyboard disclosure. The fixture imports the production store/table/drawer/CSS but uses synthetic HTTP and readings; it is not full-shell, appliance or published acceptance.

SMTP form defaults derive from the persisted type

The existing alert email editor's UIEmailConfig extends the public EmailConfig with concrete normalized password/rate-limit values. It no longer declares or defaults Reply-to, max retries or retry delay that the save/read path cannot persist. The email panel removes those ineffective controls and explains that Pulse manages retries; supported SMTP, TLS, rate-limit, routing and off-state editing remain unchanged. Destination payload builders still discard unsupported fields from stale drafts. Component, helper, destination and local browser controls cover the boundary; they do not prove SMTP delivery.

Notification recovery callbacks retire with their view

The shared delivery-health owner used by Overview and Notifications ignores pending health/action/activity completions after unmount. It cannot show a late global toast, restore retired local feedback or initiate post-action health/activity reads in a replacement view. Retired entry points do not confirm or dispatch an action. Unmount does not cancel or undo a previously issued request; active views retain their existing confirmation, latest-read ownership and persistent action-failure feedback.

Verification: useNotificationDeliveryHealth.lifecycle.test.tsx and the connected useAlertDestinationsTabState.test.tsx cover both queue actions, retired reads/callbacks and a replacement owner's independent accepted action. Desktop/phone browser controls use the real Overview/Notifications surfaces, API adapters, toast and CSS with synthetic responses, not native delivery or published queue recovery.

Existing monitoring surfaces stay truthful while mounted (10 October 2026)

Availability table order/filter/status, fleet status, open probe card and loaded Settings count/colour/label age the same retained evidence on the shared clock, not only after new props or a reload. Clock readers retire with their owners. Probe path verdicts and persisted history are not inferred from time alone. Held node, guest and Mail Gateway metric attention uses that clock; non-metric copy does not depend on its tick. Existing notification view lifetime boundaries remain.

VmwareOverview retains the page-owned workload state and renders an h2 VMs through the WorkloadsSurface tableTitle slot. Kiosk may remove the filter bar, not this name. The count is filtered rows once shared-toolbar readiness holds, including zero; no extra state, toolbar, route or hidden input is added.

Threshold Off reads as the engine reads it (11 October 2026)

The alerts feature model and the shared threshold summary helper mirror the backend's Off encoding rather than substituting their own. A backup or snapshot critical age left Off stays 0 instead of rising to the warning age, and the Docker restart-loop and memory-limit summary items read "off" for a rule or level at 0 instead of restating factory values. Per-resource rows render a muted dash for metrics an override cannot store (Docker container restart and memory-limit rules, every PMG threshold). No shared primitive, table component or layout changes.

Threshold rows hand a labelled target to the delay editor (1 November 2026)

A threshold row's delay action passes its resource ID, the alert signal from the shared getAlertResourcePreferredIntentSignal (first supported CPU, memory, disk or temperature column, else offline; desktop and mobile rows use the same helper) and a label from getAlertResourceIntentTargetLabel (the row's name and, when known, its machine or node). AlertIntentPolicyPanel lists a target that is not one of its resources, such as a machine disk or sensor child ID, as its own option under that label instead of resetting the selection to the first listed resource, and disables preview for it. It applies each handoff once, reading the resource list untracked, so an inventory change cannot replay it over the operator's later selection. The Machine Sensors section reuses the shared table with no Global Defaults row and no Alerts column; its rows stay one line, with the reading and chip in the subtitle and any readingMessage on the name's line. The threshold override read model (useAlertOverridesState) treats a note-only change as a change, so the row and its editor show the stored note after Discard or a reload. No shared primitive or layout changes.

Stopped VM filesystem presentation (12 October 2026)

getWorkloadGuestDiskRead treats an explicitly stopped VM as having no current in-guest filesystem reading, even when a power-state update has no diskStatusReason and the resource retains numeric disk fields. The same rule applies to Proxmox and linked Pulse-agent producers. Workload bars, accessible sparkline values, filesystem details, capacity sorting and the current History legend all consume that shared decision. Stored History is not removed or rewritten. On restart, an explicit prev- read deferral remains last known until fresh source evidence clears it; a real measured zero is still current.

This does not impose VM semantics on stopped LXC platform storage, alter collection, probe a guest, or establish thaw, covered-filesystem writes or native restart recovery. Regressions cover the helper, both row modes, details, History and the native-facet-to-workload mapping with an unaffected second site.

Filesystem read provenance reaches alert settings (12 October 2026)

The Alerts Guest Filesystems projection now consumes the Workloads read admission instead of treating retained numeric fields as a current reading. Explicit native stopped-VM state and unavailable reads withdraw usage; retained reads are labelled Last known. The threshold card displays a wrapping readingMessage next to the reading and the desktop row puts it on the name's line with its full text in the title, including before a missing saved mount has another sample, while preserving threshold values and override identities. Independent LXC platform storage and fresh zero remain unchanged.

Verification connects useThresholdsGuestData.test.ts, both layouts in ResourceTable.test.tsx, and the production hook/table/editor browser fixture guest-filesystem-thresholds.cjs. Presentation proof is not native guest recovery, thaw/writes/liveness, alert delivery or installed/released acceptance.

Canonical History follows the filesystem producer (12 October 2026)

On Proxmox VMs and system containers, canonical History uses the same filesystem-ownership rule as Workloads: Proxmox's filesystem list takes precedence (including an explicit empty list); the linked Pulse agent owns the fallback only when that list is absent and its own list is present. A current linked-agent reading is not deferred by an unrelated QEMU read failure, and a silent linked agent's retained aggregate is never current just because the platform keeps the resource online. Native stopped state still withdraws in-guest carriers from either producer; stopped LXC platform storage and unrelated platform metrics remain independent.

The existing aggregate is qualified, not recalculated from filesystem rows. Missing aggregates remain absent, measured zero remains usable, and neither Last seen nor a QEMU recovery renews a silent agent's report. Stored History, memory/CPU/network/IO, collection and guest commands are unchanged.

Verification: resourceDetailDrawerMetricsHistoryModel.linked-filesystems.test.ts and mounted ResourceDetailDrawer.history.test.tsx connect live/silent agents, provider precedence, stopping/recovery and zero. The local linked-guest-history.cjs browser fixture uses the production canonical drawer, Workloads snapshot adapter/drawer, History and CSS on desktop and phone, with synthetic GET responses. These controls do not establish native restart, thaw, covered-filesystem writes, installed recovery or containing software.

Guest backup completion readout (12 October 2026)

The existing guest Overview's Last completed backup row shows elapsed age (hours/minutes for recent points, days for older ones), never a rolling-day Today/Yesterday claim. A validated completion also has visible UTC date/time in a semantic time element with its original ISO timestamp. Invalid, future or absent times produce no completion element. The row still distinguishes running activity from completed protection and preserves the policy's tone; the passive clock may rejudge a previously future timestamp when it becomes valid, but cannot supply new completion evidence.

GuestDrawer.backup-clock.test.tsx and the connected drawer/model controls cover these states and update on time alone. Desktop Chromium and touch-phone WebKit exercise the real policy producer, consuming drawer, shared clock and existing table evidence. No route, guest check, backup, threshold, collector or permission changes. This presentation result does not establish thaw, covered filesystem writes, native restart recovery or containing released software.

Client backup report scheduling and data-quality states

Reports adds Client backup health to the existing schedule editor. PDF/CSV, weekly/monthly cadence, current workload/tag scope and existing email/disk delivery controls reuse the schedule workflow. Empty selection and tags mean all current VMs and system containers. The picker restricts this kind to those workload types and has no performance-report selection cap. Its existing default cap remains unchanged for other consumers. Switching kinds retains only valid workload selections and preserves the existing digest/inventory constraints.

The schedule card presents last report data quality separately from delivery acceptance. Missing/incomplete history never reads as zero failures or completed delivery failure. The browser payload cannot supply server-owned quality fields. Copy states the PVE task scope, PBS destinations and unsupported-provider state. Pricing and current advanced-reporting/branding entitlement behavior are intact.

Proofs: ReportingPanel.test.tsx, ResourcePicker.test.tsx and reportingSchedulesModel.backupHealth.test.ts under Settings tests cover report kinds, quality/delivery coexistence, unrestricted workload selection, payloads and kind switching. Final-source browser proof covers the actual editor, client scope, reload, delivery/recovery states and narrow/desktop layouts.

The shared picker uses semantic theme colours for workload names, selection, tags and chips. Large selections remain in a bounded scrollable summary so client filters and delivery controls stay reachable. Toggle rows expose pressed state, and keyboard removal scrolls selected chips into view. The browser matrix covers light and dark themes at all three widths.

The additive reporting-catalog backupHealthAvailable flag describes runtime renderer presence independently of entitlement. New backup schedules are offered only when it is present, while existing schedules remain editable after a runtime change. Legacy catalogs default to false. API and catalog-model tests cover absence and registration so staggered source upgrades do not offer unavailable rendering. No price or entitlement changes.

One explicit guest backup completion time (12 October 2026)

The Workloads backup badge and name-cell indicator disclose the same UTC completion date/time as the guest drawer, including on the fine-pointer hover shortcut. formatBackupCompletionTime normalises the original instant once per disclosed readout; the semantic time preserves its ISO timestamp. Local browser timezone and locale cannot silently change the visible calendar day or time. The completion text wraps at narrow widths. Unknown/future/absent readings still have no completion element; running or ended activity and clock ticks do not renew it. Only replacement evidence changes the supplied completed time.

BackupCompletionTime.test.tsx, BackupStatusAccess.test.tsx and the formatter controls connect both disclosure paths to the drawer, offset inputs, updates, uncertainty and disposal. The backup-completion-time browser fixture uses real VM/LXC rows, policy producer, hover, keyboard/touch dialogs and the guest drawer in non-UTC desktop and phone contexts. This is unambiguous presentation, not PBS verification ingestion, independent thaw, filesystem liveness, native guest restart recovery or containing software availability.

Host drawers distinguish unknown filesystem usage from empty filesystems

The existing Proxmox node and Docker host per-filesystem rows use the shared host disk presentation. An explicit negative usage sentinel, negative or non-finite used bytes, or absent/non-positive/non-finite capacity produces Usage unavailable, with any finite positive capacity retained separately. Unavailable usage has neutral text and no utilization bar; a retained capacity alone is not a healthy zero. Finite byte ratios preserve their existing threshold presentation. Real zero-byte values, including zeros omitted by the canonical DiskInfo JSON representation, remain valid empty filesystems when capacity is known and no unknown sentinel is present.

The shared usage text wraps in both existing drawers, rather than relying on a tooltip or rendering an old percentage after a new unavailable read. The existing mount filtering, row disclosure and single-filesystem summary stay unchanged. The connected HostDrawer.filesystem-readings.test.tsx and shared DrawerDiskListCard.readings.test.tsx controls cover unknown/capacity/zero, invalid bytes and same-mounted-drawer replacement. The desktop Chromium and narrow touch WebKit browser-tests/host-disk-usage.cjs fixture consumes synthetic Resources responses through the actual platform tables and drawers. No new route, monitoring policy, provider operation or native recovery claim is added.

The shared filesystemByteUsage.ts admission also applies to filesystem inputs of stackedDiskBarModel.ts: unusable byte rows remain neutral in tooltips and do not contribute used bytes, capacity, bars, combined percentages or highest usage. All-unmeasured filesystem lists with no independent accepted aggregate render N/A, not a fabricated zero. Independently supplied aggregate metrics and retained/unavailable source classifications preserve their existing semantics. The connected stacked-model controls retain explicit aggregate fallback tests; the browser fixture checks mixed and all-unknown lists in the actual summaries as well as both drawers, including live replacement at narrow widths.

The aggregate bar admits byte sublabels only with finite non-negative measured used bytes and finite positive capacity. An independently valid percentage remains visible beside invalid capacity without formatting NaN/Infinity bytes. Model/component controls cover non-finite, absent and non-positive capacity; workload-disk-metrics.cjs checks both Bars (including the tooltip) and Sparklines through the same REST reading at desktop and phone widths.

The storage pool drawer left the form-select-shell consumer list: its usage chart now takes its range from the shared drawer history toolbar (GuestDrawerHistory), the same control as the guest and node drawers, rather than a pool-local FormSelect. The guest drawer's embedded disk detail keeps its FormSelect. Guardrails that pinned the retired pool card classes now pin the drawer pieces the pool Overview composes (attention section, Worth checking, drawer history, closed Details of DetailSectionTable sections).

features/platformPage/RecordOwnerLink.tsx is the one owner link for record drawers: an InlineDetailPanel actions button, styled as a drawer header action, that opens the object a record is about through buildResourcePagePath. It renders only when that path carries a parameter the target page reads to open or filter to the object (buildRecordOwnerPath), so a record never offers a link that lands on a whole tab. Platform record drawers compose it instead of hand-built anchors, and every record drawer is an InlineDetailPanel inside InlineDetailTableRow: the Docker Swarm tables share one panel in DockerNativeTableShared.tsx, Kubernetes records share KubernetesRecordDetail.tsx, and hand-written dl/grid detail rows are retired from Docker networks and storage usage, Proxmox replication, coverage and recoverable points, and Agent Doctor.

Full-width helper rows span the visible columns, never a guess. A colspan={100} windowing spacer or colspan={99} storage group row created phantom fixed-layout columns, and on phones (where hidden columns leave the visible percentages short of 100%) those columns took the rest of the table: rows and drawers on windowed Kubernetes tables stopped at 70% of the width. components/shared/tableVisibleColumnSpan.ts counts the header's visible cells (re-measuring on resize and header changes), and PlatformWindowedRows spacers, StorageGroupRow and the storage pool spacers use it. DrawerAttentionSection also owns DrawerNoAlertsLine, the one "No active alerts" line every drawer shows (with the optional availability note), and the attention card's subject wraps instead of truncating. features/platformPage/platformDetailTimestampRow.ts is the one drawer row for a reported timestamp (absolute time and age, wrapping, an optional sentinel year), replacing three copies.

Platform tables that a page drives from the URL pass the page's search through createPlatformTableFilterState's externalSearch/onExternalSearchChange rather than a table-local signal: the TrueNAS Storage, Apps, VMs, Shares, Services and Protection tables now do this for ?q=, as the Kubernetes tables already did, so RecordOwnerLink can land on a narrowed tab whose own search box shows and clears the filter. ArtifactSourceBadge takes the opt-in MetadataBadge wrap for drawer use; table rows keep the one-line badge.

.table-fixed td and th (index.css, @layer utilities) set nowrap, hidden overflow and an ellipsis, and the class-plus-element selector outranks a single utility class such as whitespace-normal on the cell itself, so that class on a fixed-table cell does nothing (the rule also reaches cells of a table nested in a drawer row). A value that must wrap sets the wrap on its own element inside the cell, as DetailSectionTable does with block whitespace-normal; the Coverage drawer's restore evidence Location and Details values now follow that pattern.

PlatformHostScope.tsx owns the host-to-workload scope interaction for platform overviews: PlatformHostScopeNameButton (the host name as a pressed toggle naming the workload section it controls), PlatformHostScopedTableTitle (" on " with its count and the clear button) and createPlatformHostScopeFocus (selecting a host focuses and reveals that heading). The Proxmox, Docker and vSphere overviews use them, so the three stay one shape. A host table that scopes keeps its details chevron on phones and takes the platform-table-mobile-w-38 identity column (Uptime drops to 10%), the same allowance the Proxmox node column has for the toggle and dot before the name.

The browser Assistant chat window was retired on 2026-10-22, taking its drawer, launcher, command-palette entries, chat action-icon and copy shells, and the Assistant hand-off buttons on alerts, Patrol, and resource drawers with it. frontend-modern/src/stores/aiAvailability.ts now holds the shell's AI-enabled fact that the chat store used to carry, and the Patrol tool transcript block moved to frontend-modern/src/components/patrol/ with its CopyValueButton composition intact. The External agents settings page lives at /settings/pulse-intelligence/agents; the old /assistant path redirects there.

Source-row error banners follow the row's severity

InfrastructureSourceManager.tsx renders a row's error banner, in both the table and the narrow card layout, from one rowErrorBannerClass helper keyed on row.lastErrorTone. useConnectionsLedger.ts sets that tone to warning for a partial-read error on a degraded connection and to critical for every other error, so a reachable source with one failed reading gets an amber banner beside its amber "Degraded" badge, while outages keep the rose banner. No new component, primitive or badge vocabulary is added. The rendered tones are pinned in InfrastructureWorkspace.test.tsx.

Configuration transfers retain their confirmed workspace

Recovery's overview identifies the selected-workspace configuration export as limited rather than a full-server backup before either dialog is opened. It names Proxmox connections, discloses missing history, TrueNAS/vSphere, local login credentials and agent enrolment state, and links the shipped migration guide for complete scope and full-state recovery. Restore warns to back up the destination and describes replacement of included settings and API-token records, never a merge option. RecoverySettingsPanel.test.tsx covers the visible scope, and existing dialog actions; desktop/narrow browser proof must inspect the overview and follow both actions into the production dialogs. Archive formats, server application and access authority do not change.

All login configurations use the same backup-passphrase form. Export asks for a strong, unique passphrase of at least 12 characters, explicitly different from Pulse or SSO login credentials, and a second masked confirmation. This encrypts an archive; it is not a sign-in check. Neither the recovery overview nor Settings wiring selects a login-password mode. Both fields use new-password autofill semantics and explain separate storage and the lack of forgotten-passphrase recovery. The transfer flow rejects missing, short or nonmatching values before credential prompting or dispatch, preserves them for deliberate correction, and sends only the original, unchanged passphrase, never the confirmation. Pending controls remain locked; successful export, close and switching dialogs clear both fields. Import still accepts the original password/passphrase for old archives, not a newly changed login credential, with no confirmation or new minimum-length requirement.

The transfer flow/dialog tests cover these validation, exact-byte, auth-mode, pending, explicit-correction and cleanup behaviours. Settings architecture and overview tests cover the removed auth-dependent wiring. The shipped migration guide matches the form and preserves older-archive guidance; its rendered access tests check both. Desktop/narrow browser proof must exercise typo rejection, explicit correction, an actual synthetic export download, original-passphrase import and the migration guide through their production consumers. This changes neither backend encryption/authorization nor existing archive formats.

The overview has its own settings-recovery-overview verification policy, requiring RecoverySettingsPanel.test.tsx. That copy/action test is not transfer safety proof: settings-configuration-transfer retains its existing dialog/flow test alternatives without the overview test. Changes spanning both policies must satisfy both; generic settings framing cannot substitute for either.

Recovery's export/import dialogs capture and display the effective workspace ID when opened, including explicit default scope. Their production flow freezes inputs before asynchronous file reads, binds both POSTs through the API client's expectedOrgID and retry: false, and rejects a changed workspace before dispatch (including the client's post-CSRF recheck). An invalid organisation must never fall back to default, and a CSRF refusal must not replay a transfer. Only one transfer can be pending; controls and dismissal remain disabled while waiting.

Token entry saves a credential without submitting either transfer. The operator must review the target and submit explicitly, including after a server refusal. A failed/unknown import response is not proof that nothing applied: persistent in-dialog guidance requires checking the destination before another attempt. Transfer errors and token-saved status belong to the active dialog, never a duplicate foreground toast on the inert notification portal. Token validation appears in the nested token prompt; closing it returns the retained import warning and explicit-resubmission status to the transfer dialog. Dialog content scrolls independently so feedback, fields and actions remain reachable on a short phone screen. An import completed for a workspace that is no longer selected keeps its original destination and no-reload outcome visible until the operator dismisses it, with the submitted credentials and file cleared. Exports discard late data after a workspace change; successful imports identify the original workspace and never reload a different one. Leaving Settings aborts the client's wait and suppresses late downloads, notices and reload timers; it does not promise to undo a server-side import or prevent transport-level replay. A non-retrying application fetch is not server-side single-application protection. Existing archive formats, authorisation, scope and encrypted payloads remain unchanged.

useBackupTransferFlow.test.ts runs the real API client across delayed file/CSRF preparation, scope switches, duplicate calls, non-retryable failures, token entry, unmount, archive compatibility and cleanup. BackupTransferDialogs.test.tsx checks target/error disclosure and pending input/action states. Browser evidence must exercise the production flow and consuming dialogs at desktop and narrow widths with safe responses, not a replacement transfer implementation.

Backup evidence retains its measured native scroll extent

Proxmox Coverage and By date use the shared PlatformWindowedRows renderer, not a fixed-height backup-specific runway. An expanded evidence panel belongs to its keyed summary item: scrolling inside it must not unmount that item or replace the panel with unrelated workload rows. Group headings count as their own items; measured offscreen panels retain their scroll extent. Live inventory replacements retain the mounted row, its disclosure focus and its expanded state. Spacer cells span only the currently visible columns, while the shared desktop/phone mounted-item budgets remain bounded.

ProxmoxCoverageTable.test.tsx exercises both consuming tables, grouped dates, desktop/phone budgets, tall panels, snapshot replacement, scroll-away and return, and spacer spans. Browser proof must scroll actual production evidence panels at desktop and narrow widths and inspect pixels. This changes only presentation: posture, identity matching, history, permissions and native backup/guest/recovery operations remain unchanged.

Backup table extent survives measured-detail retirement

Coverage and By date opt into PlatformWindowedRows.reserveScrollExtent. The shared measurement owner reserves the natural table height on its outer scroll shell before a tall inline panel leaves the row window. This prevents a transient shrink from clamping native page or app-container scrolling. It does not set the table's own height or move the native scroll position. Natural remeasurement releases the reserve on disclosure close or responsive layout; leaving windowed mode and unmount restore the shell's previous minimum height. Other table consumers keep their existing lifecycle by default.

PlatformWindowedRows.test.tsx checks opt-in extent retention, bounded rows, return, close and small-result filtering; consuming backup proof must exercise real desktop/narrow app scrolling and inventory polling, including the last visible workload. Posture, keys, authority and native recovery are unchanged.

The reserve derives from the full keyed measured-height map plus table chrome, not a DOM height sampled while a window retires. Offscreen detail measurements remain in that map until the owning item is remeasured or removed, so a transient DOM extent cannot lower the reserve before its compensating spacer settles.

Backup activity distinguishes unknown size from absent restore points

The By date activity view keeps completed, in-window entries even when their sizes are unavailable. Count and day selection remain usable in Volume mode. Volume bars and axes measure known positive finite sizes only; unknown sizes have a separate amber day marker, explicit notice, day label and source tooltip. Mixed totals are labelled as known volume, never a complete total, and source percentages are omitted when any size for that day is unknown. An all-unknown window has no numeric volume scale. Empty days remain distinct from unknown size.

The existing source, search, node, type, location and local-calendar window scope applies to both known and unknown sizes. Running/failed artifacts and unknown or future dates do not become completed activity. Posture, backup completion, permissions, guest operations and actual recovery are unchanged.

ProxmoxBackupsTable.test.tsx exercises unknown-only and mixed inventories, Count/Volume switching, selection and source scope through the HTTP consumer. proxmoxBackupActivityPresentation.test.ts checks separate unknown counts, excluded dates, labels and per-source tooltips. Browser verification must inspect the real table/chart with safe inventory at desktop and narrow viewports.

Coverage task order is evidence, not an input-order tie-break

The Coverage task summary names a latest task only when the observed start times establish one (or only one task was observed). Missing, invalid, non-positive or future starts cannot be ranked as epoch dates; distinct observations with equal latest starts cannot be ordered by input position, status or ID. In either ambiguous case the summary is Unknown, not OK or None. The existing detail explains the uncertainty and retains every possible latest task's status, start, duration, literal error and full ID. Possible tasks are displayed in a stable non-time order; older known tasks are not possible latest tasks.

The task column remains visible even when no latest task can be established. Existing workload/search scope includes candidate errors and IDs, never another connection's namesake. Ordinary inventory refresh can restore an ordered latest, reintroduce uncertainty or remove all task evidence without replacing the keyed row, open panel or disclosure focus. No new API read, route, association rule, protection assessment, backup/guest operation or recovery guarantee is added.

proxmoxBackupTaskChronology.test.ts checks permutations, ties, future/unknown starts, candidate completeness, scope and sole/absent tasks. ProxmoxCoverageTable.test.tsx checks the canonical compact/full consuming summary and complete candidate detail. backupTaskEvidence.test.tsx exercises the parsed HTTP inventory through the router, compact/full summaries, complete literal evidence, saved-view search and ordinary polling. Browser verification must exercise these states through the production consumer at desktop and narrow widths and inspect actual pixels.

Unmatched backup placement comes only from node evidence

Unmatched PBS namespaces and datastores remain backup locations, not guest nodes. Unmatched PVE records retain their reported native node, never their connection name as a substitute. Both existing workload detail views say Not reported for missing VM/LXC placement; host labels remain separate. Node filters and unscoped task matching cannot use invented fallback nodes.

proxmoxBackupRecoveryModel.identity.test.ts and the normal-suite backupWorkloadIdentity.test.tsx cover absent/blank placement, independent reader/location identity, native-node controls, node filtering and live owner arrival/removal through the parsed inventory consumer. Matching rules, source keys, restore points, protection policy and guest operations are unchanged; removing fabricated node metadata prevents one false unscoped-task match. Desktop and phone browser proof exercises both existing detail views. This corrects attribution, not #2681/#3248 causality or native recovery.

When Coverage hides the Location column on narrow containers, each existing restore-evidence Source cell also wraps its labelled repository location. Removing a guessed node must not discard the useful namespace/datastore it misrepresented. ProxmoxCoverageTable.test.tsx checks narrow/full projection and complete long location text; the parsed inventory regressions also retain reader identity and full repository context without adding reads or routes.

Unread backup storage is named once on the Backups page

When the postures on the page carry unread provider scopes, the Backups page shows one amber status line under the health strip that names them (up to three, then a count) and says Pulse cannot confirm the backups stored there, with one action: check that the storage is reachable, or check the API token's permissions when every named read was rejected. It reuses the existing read-notice style, adds no request, and the connection name appears only when the gap spans several. The per-workload drawer keeps the server's explanation, which names the same scopes. proxmoxBackupUnreadSources.test.ts and ProxmoxBackupsTable.test.tsx cover it.

Backup rows stay on one line (1 November 2026)

Coverage and By date rows keep one line at every container width, phones included: the workload name truncates beside its folded type/target ID, and a PVE or reader connection is named inline only when the loaded inventory spans more than one, where it tells namesakes apart. The page decides that over the whole inventory, so labels do not appear and vanish as filters narrow the list. Narrow By date rows are workload, age and state; source kind, reader connection, repository location, size and provider identifiers live in the single-point drawer, whose disclosure name already carries both scopes before it opens. This supersedes the earlier wrapped-identity projections below 720px, which grew phone date rows to 116-132px and Coverage rows to 65-81px and pushed the list below the fold; measured on the mock rig the rows are now 37px and 32px (desktop 47px to 33px). Matching, posture, verification, filters, polling and the API are unchanged.

backupWorkloadIdentity.test.tsx, recoverableIdentityAccess.test.tsx and backupSourceConnection.test.tsx exercise the parsed inventory consumers at 320-1200px: single-line truncation, inline connections only for multi-connection inventories, drawer identity and live refresh. Browser proof measures row heights and inspects actual pixels at desktop and phone widths.

Location-filtered backup evidence is not a location protection rating

The Proxmox backup-health strip names its all-target/all-location count scope, independent of filters. In Coverage, a resolved backup-location filter qualifies ages and restore evidence next to the list and inside an expanded record. The server's protection assessment, provider evidence and task history remain workload-wide; a current backup elsewhere must not read as protection of a stale, running-only or snapshot-only selected location. Labels come only from resolved inventory identity, never raw saved-query text. An unresolved location keeps the existing explicit empty selection rather than broadening to another repository. Clearing the filter removes the qualification without replacing the keyed open record or making another read. No assessment, association, policy, API, route, guest operation or recovery authority changes.

Normal-suite ProxmoxBackupsTable.test.tsx covers stale, running and snapshot-only selected evidence beside a server-owned Protected workload, all-target counts, independent task history, filter clearing and ordinary-poll disappearance/return. Desktop/narrow browser proof must exercise the parsed inventory consumer, filter interactions, scoped open records and literal long labels and inspect actual pixels. This corrects evidence scope, not original incident causality or native backup/guest recovery.

Backup activity preserves calendar-day interaction identity

BackupActivityChart retains each day control by its local calendar key across inventory refreshes, scope changes and metric changes. The focused control must not disappear or silently become another date when new buckets replace the input. A removed focused day returns to the selected activity-range control, without changing the saved day filter or moving focus from another control.

The open tooltip reads the current bucket for its day, not a retained count or volume snapshot, and follows the actual control after inventory layout changes. Removed days withdraw the reading, including on return until a new interaction. Unknown-volume markers and counts update together. Reconciled chart values must not mutate the caller's inventory-derived DTOs.

Normal BackupActivityChart.test.tsx covers focus, rollover, disappearance, current tooltip readings and input ownership. ProxmoxBackupsTable.test.tsx exercises the parsed inventory and ordinary polling through the consuming date filter, preserving bounded reads. Backup/posture policy, source association, read cadence and guest operations are unchanged; this is not native recovery.

PlatformResourceCounter can name what it does not count (3 November 2026)

PlatformResourceCounter takes an optional detail string, shown after the count inside the same live status ("2982 restore points · 33 running"). It is for rows a table lists but its noun does not cover, so the count can stay true to the noun. Callers without it are unchanged.

Backup polling does not choose a different workspace

Bare legacy Backups links retain the useful initial health-based view choice, then bind it to the existing canonical view route once inventory is readable and the first assessment settles. History replacement preserves shared query filters and the fragment without a scroll jump. Explicit named/query views remain authoritative. Subsequent health changes update the mounted workspace, not its selection; existing tab and health-count actions remain the way to choose another view. This adds no route, assessment or backend mutation. ProxmoxBackupsTable.test.tsx verifies focused open evidence across both polling directions, deferred/failing assessment, retry and URL/history compatibility.

Compact disk summaries preserve producer aggregate ownership

In aggregate mode, a usable supplied aggregate owns the total percentage and byte summary even when filesystem rows are present. Aliases in the breakdown must not be counted as independent capacity. Mount rows retain their own values, severity, order and tooltips; explicit max and per-filesystem modes are unchanged. Without a usable aggregate, measured-row fallback remains. Independent percentages must not borrow byte totals from the breakdown. Current/last-known/unavailable classification and unknown-filesystem admission remain authoritative.

StackedDiskBar.test.tsx exercises alias/reorder/zero, invalid and percentage-only aggregates, fallback, severity, freshness and mounted replacements. DisksCard.test.tsx checks that its normalized ratio crosses the aggregate component boundary as a percentage. Actual producer frames must pass through canonical resource and Workloads projection into desktop/narrow consumers, with pixels inspected. This is presentation correction, not native guest recovery.

Coverage distinguishes assessment clocks from visible restore evidence

An expanded Proxmox Coverage record exposes the server assessment's evaluatedAt, lastSuccessfulPointAt and lastVerifiedAt as separately labelled UTC instants. Time labels and values use wrapping stacked rows, not truncated hover-only labels. Evaluation is not backup completion. These workload-wide recorded times do not come from the current artifact list or its location filters and do not establish a fresh inventory read, successful restore or guest responsiveness. Missing optional clocks say Not reported; malformed instants say Unknown rather than being repaired by the browser or borrowed from another field. A failed posture read withdraws the section with the assessment; ordinary recovery restores it. No provider, policy, association, age, permission, guest operation or polling contract changes.

Normal-suite coverageAssessmentEvidence.test.tsx exercises the parsed Backups consumer at desktop/narrow widths, differing artifact/evaluation/point clocks, missing fields, location filtering, ordinary refresh, failed-read withdrawal and isolated retry. ProxmoxCoverageTable.test.tsx protects strict presentation of invalid instants. Browser proof must inspect the actual consuming record at desktop and narrow widths, including UTC text, keyboard/touch disclosure and refresh/recovery with the open record preserved. This is evidence inspection, not original incident recovery.

TrueNAS memory composition is inspectable without hover

The existing TrueNAS system memory bar opts in to a resource-named native button and the shared accessible Dialog. Keyboard and touch expose the same live stackedMemoryBarModel rows as pointer hover, not a second calculation or a snapshot captured when opening. A measured zero stays zero, percentage-only readings do not acquire byte/free/cache rows, and absent ARC never gains a fabricated cache row. Metric withdrawal removes its open breakdown. The native button stops the enclosing system-row action; closing/Escape/backdrop returns focus through the existing dialog owner. Its 44px phone target and wrapping labels must fit narrow viewports. Shared hover suppression remains unchanged; other consumers do not opt in automatically. No collector, freshness, alert, permission, API call, polling cadence or native recovery guarantee changes.

Normal-suite TrueNASSystemMemory.test.tsx uses the actual table, bar, model and dialog for known/missing/zero cache, measured zero, percent-only/unreported readings, live resource changes, row isolation and unavailable withdrawal. StackedMemoryBar.test.tsx retains unavailable explanations without fabricated usage and Dialog.test.tsx owns modal isolation/focus. Browser proof exercises the consuming TrueNAS table with ordinary keyboard/phone taps, hover parity, updates and dismissal, and natively inspected desktop/narrow pixels. This repairs composition access, not the reporter's missing native ARC or I/O/History source.

SSO endpoint cards remain readable and copyable on narrow screens

The SSO provider card in SSOProvidersPanel.tsx separates each endpoint's label from its value with a definition-list entry. OIDC callbacks and enabled SAML metadata/ACS values wrap anywhere rather than truncating or forcing an unshrinkable inline URL row on the card. The shared CopyValueButton remains a native, keyboard-operable control with a preserved copy icon, a full-width 44px-tall narrow target, and the exact server-supplied value. Copying does not navigate, normalise or infer a URL. Existing missing-endpoint guidance, disabled-provider visibility and read-only management gates remain; viewing/copying is not a settings mutation or a paid SSO gate. SSOProvidersPanel.test.tsx verifies exact OIDC/SAML clipboard values, read-only/disabled-provider controls and clipboard rejection. Actual consuming desktop and narrow-browser geometry, keyboard/touch and inspected pixels establish responsive fit beyond JSDOM.

Failed login discovery is not first-run setup

Login.tsx uses a complete bootstrap observation or reads /api/security/status before selecting setup or sign-in options. A boolean hasAuth hint alone cannot supply SSO/local-login policy. A failed read or missing/non-boolean hasAuthentication remains unknown; it never becomes false. Only an explicit server false enters the existing bootstrap-token-protected setup wizard.

The failed-read surface offers sign-in help and one explicit, read-only retry. While pending, its live status and retry control remain mounted, duplicate requests are ignored, and a successful retry hands owned focus to the access region without stealing focus the user moved elsewhere. Disposal aborts the read and ignores late results. Supplied valid discovery still avoids a redundant read; existing SSO hiding, local-login override, lockouts, demo sign-in and remembered username behaviour remain unchanged. No authentication, entitlement, credential, server setup or session-validation rule changes.

Normal-suite components/__tests__/Login.test.tsx covers failed and malformed reads, recovery to SSO/local login or confirmed setup, single-flight/cancellation, focus ownership and the established sign-in behaviours. Browser proof exercises the actual Login/API client/CSS with safe HTTP fixtures at desktop and narrow widths; it is not installed SSO or original private/access recovery.

Replication inventory failure presentation (30 October 2026)

The Proxmox Replication table consumes the validated full inventory owned by unified-resources. Malformed successful HTTP responses use the existing PlatformErrorState and Refresh interaction with a fixed safe explanation: missing jobs do not establish that none are configured. Neither raw provider body nor JSON parser details enter the rendered error. This is a failed read, not an empty-state or a new error surface; full retry, normal status filters, phone columns and disclosure interactions retain their existing owners.

By date retains literal restore-point date evidence

The existing Proxmox By date record exposes the provider's exact reported date beside its formatted Created interpretation. Offsets and subsecond precision remain literal wrapping text for PBS snapshots, PVE backup files and guest snapshots, including malformed and future values; missing values say Not reported. Unknown Created values remain unknown. The existing By date notice continues to explain unavailable/future dates and their exclusion from activity and day filters; no duplicate per-point warning is added. Raw dates are not an inferred completion time, freshness, protection or restore assertion.

The same open point updates its raw date and interpretation on ordinary inventory polling without borrowing another point's date or losing owned focus. Normal-suite recoverableIdentityAccess.test.tsx covers all three sources, precise/malformed/future/missing/literal-markup dates, desktop/narrow widths and in-place replacement/withdrawal. Registered ProxmoxBackupsTable.test.tsx protects unknown-date evidence through the parsed inventory consumer and ProxmoxRecoverableTable.test.tsx protects direct presentation. Browser proof must exercise the actual By date consumer at desktop and narrow widths and inspect wrapping, uncertainty, keyboard/touch disclosure and live replacement. No parser, association, ordering, count, filter, posture, API, permission or provider-operation contract changes. This is inspection, not guest recovery.

Polling can move an existing By date point between calendar groups, including into Unknown date. Before the keyed list changes, the table captures only focus it owns. After the move it restores that same connected control only if focus fell to the document body; it never steals focus moved elsewhere or resurrects a removed point. The open record, strict dates and existing withdrawal fallback remain unchanged. Direct normal-suite tests cover owned/external/removed focus; the parsed Backups consumer covers an open point moving into Unknown date. Actual Chromium and WebKit browser proof must cover that day-group move.

The Proxmox gzip baseline increases by 256 bytes (80339 to 80595, 0.32%) for literal date evidence and the browser-reproduced focus repair. The 15% threshold, all other per-chunk baselines, total JS/CSS and preload/SRI checks stay unchanged. This is a narrowly declared intentional UI/accessibility cost, not suppression of the retained failed budget runs or permission to omit browser verification.

Plans & Billing MSP referral copy stays in the billing presentation owner

The Pulse MSP line at the end of the Plans & Billing Plan tab takes its text and button label from SELF_HOSTED_PRO_BILLING_PRESENTATION (mspReferralBody, mspReferralActionLabel) and its tagged destination from PULSE_MSP_REFERRAL_URL in selfHostedBillingPresentation.ts, like the rest of the billing shell copy. ProLicensePanel.tsx renders it with the shared ButtonLink outline action, which supplies noopener noreferrer for the new tab; no local link styling or copy is drafted in the panel.