Skip to content

The ultimate harness for coding agents and beyond. All your agents. All your machines. One command center. Start with code, then follow your curiosity and build across disciplines: CAD, circuits, robots, games and music.

63.5 KB
Raw Blame

Windows dial end-to-end report

Status: E5.2 COMPLETE WITH LIMITS; CLEANUP WAIVED BY OWNER on 2026-10-20. Report updated 2026-10-20T14:59:04Z. Code implementation and whole/scoped review are complete, with I1/M1 resolved. All eleven scenario verdicts are recorded with their limits; missing physical/user-switch/unplugged checks remain NOT RUN, S7 remains partial, and S11 SSH uninstall remains FAIL/partial.

All E-owned runtime processes/tasks were stopped. The owner waived the failed account/path cleanup and directed no further cleanup, reboot, unload or deletion attempts. Two test profiles, test keys, portable E root, shared scratch and positive ownership receipts remain. Final baseline comparison remains NOT PASS with six unexplained shared-state differences; no global preservation or product uninstall PASS is claimed. Cleanup is no longer a completion blocker under the owner's decision. Root will close E5.3 after documentation commit and review.

The acceptance target is the Harness desktop app, E's CLI and native serial helper together on the x64 Windows test PC with a real round Harness device (called the dial below). Arm64 hardware acceptance is outside this report.

Source and build provenance

The owner authorized integrating C into E for combined desktop/device testing. The integration reviewed and locally validated was 3426d9c8511469e555da7903f248272afbf3f3ba, combining E parent d67bd9698feb52b3d2b1e6933d08674aa859af73 and C parent 87779178ea5d119dff43d7bea6f03c1dd76fd02b. The original combined hardware-run source is 4c3def3cccaf0e820de63aeeb742a9852c6ce95a. The core coverage correction has passed its required gate. All four jobs of the exact-source native run and both architectures of the combined desktop build succeeded. The isolated portable profile ran in the owner account's visible Windows session and has completed its normal QR login. The original hardware receipts, including the S8 failure, use product input 4c3def3c; their local head was 030ba49cf after test-kit fix 23bba54ed and shipping documentation. The first completion-hook change c95e86208dafb484ee9a662972b7a196bffaa6d3 failed its decisive hardware rerun. Round 2 4fe12de12d7e1d8695bae0b01529a432fec28a17 passed native bundle-entry, CLI3 selection and decisive normal-CLI recovery. Current repository HEAD is 444139f69e9aa8b2dd92503702bb8bfedf35c99e, the eight-path final I1/M1 fix wave; its scoped rereview PASS resolved I1/M1 with no new findings. No new GUI binary or hardware result is associated with that final wave. Current private CLI selection is 4fe12de12; app/native remain 4c3def3c, and the separately hashed PowerShell flasher is unchanged. The private 0.9.0-dev.dial2 upgrade completed at 2026-10-20T11:16:27.2047485Z; native checks verified configuration/current version/helper pins and an unchanged before/after persisted-verdict file hash. The earlier dev2 hardware run had mixed provenance: app/native artifacts remain 4c3def3c, while CLI/PowerShell inputs are c95e86208. No receipt establishes that the whole runtime is c95e86208. A flash/reattachment observer passed, but its baseline dial was already attached; it does not decisively rerun the original negative-verdict recovery failure. A later genuine mid-write cancellation and direct recovery rerun did reproduce the failure on this mixed c95e86208 CLI/PowerShell runtime (scenario 8).

E5 input / candidate Recorded value
Original hardware-run source commit and tree Commit 4c3def3cccaf0e820de63aeeb742a9852c6ce95a; tree 2671233afc41a1485a56e4ce41f146f8f569ddd5
Tested round-2 CLI source 4fe12de12d7e1d8695bae0b01529a432fec28a17; independent scoped review PASS, local selected suite 103/103 + typecheck PASS; CLI3 selected with verdict bytes preserved; native entry fixture and decisive normal-CLI hardware recovery PASS
Final I1/M1 source / native PS1 fixture 444139f69e9aa8b2dd92503702bb8bfedf35c99e; scoped rereview PASS (I1/M1 resolved, no new findings); native PowerShell 5.1 function-only fixtures 7/7 PASS, no new GUI hardware run
Final PS1 source checksum SHA-256 07d51c80b43799aeef0948677ff3f14a0015c4e6b9576577e8e7ba502998aad8; actual hardware receipts still use earlier c95e86208 PS1
Tested CLI3 bundle checksum SHA-256 8cacd1f7cb52228cf356d47ba7e2cda90530cb1ab93901af6d85a87ad856199b
Current tested mixed inputs CLI 0.9.0-dev.dial3 / 4fe12de12, PowerShell c95e86208, app/native 4c3def3c; decisive normal-CLI recovery PASS; earlier dev2 hardware failure retained, no new dev3 GUI flash
Historical dev2 CLI bundle checksum SHA-256 af87a5d0e61f23f2be2ee3dd2528eab41dbe63aa883721d4985748646104e6a5
Recovery PowerShell flasher checksum SHA-256 681b96cea559a301189223ed629b15f9b729932ea3f85cc00e8dc1db9da6847f
First recovery-change native workflow (historical validation) 39412223533, exact c95e86208 source; all four jobs SUCCESS; each architecture: contract tests 739 passed/56 skipped, flasher 31/31 plus official 0.0.92 no-flash, exact watch install and clean uninstall
Relevant fix commits since the integration above ddee034120a436d742c4c5bd255ea9bdd7444989: PowerShell 5.1 fixture; 4c3def3cccaf0e820de63aeeb742a9852c6ce95a: device-service coverage correction
Original windows.yml run URL and native x64/arm64 results 39390890060, exact final source; all four jobs SUCCESS, including x64/ARM64 watch startup and clean uninstall
x64 harness-pty-windows-x64 artifact ID and digest 12250700212; SHA-256 ac3a85be12c52ad234a28ea00f41114bc952f1d929a9d3a64b615d32ae23bd1b
Installed harness-serial.exe checksum SHA-256 d6e881453d4215ba404804c05ab1f5714b96c5ab716a3c073b6fed6b1f2af11c, exact artifact match; selected PTY version 0.1.0
CLI development version and bundle 0.9.0-dev.dial1; bundle from 4c3def3c PASS in 15.7 s; installed version confirmed
Local install manifest checksum SHA-256 7cc3983d536e76fac804c6e620948d3705d8e56986db0175364fd8cea53ffd83
Local install zip checksum SHA-256 4bef2f380471f3d11e65f104a505e79930766b33369835b26060871d0752b402
Combined Windows desktop run, artifact, version and checksum 39393250390, both architectures SUCCESS; x64 zip 1.0.4-dev.dial1, 6,187 entries, SHA-256 5c2491e8197b310b809be2f744badb2787f4792b72600e32e4fa957139c4f4b6
Windows and PowerShell versions Windows 10.0.26200.0; native Windows PowerShell 5.1.26100.9711
Starting firmware, older test firmware and target firmware versions Started 0.0.92; actual CLI flash to older 0.0.91 and cable update back to 0.0.92 completed (Script PASS). Older selected from read-only official listing (92 versions); official HTTPS latest metadata: 3,360,672 bytes, SHA-256 eadb3846a6658026048418804a60341adc45fdb24793ebc9cabd782def0c8458
Redacted CLI install evidence Setup record: installer exit 0, version and installed helper confirmed
Combined portable app and signed-in workspace evidence Exact staged executable visible in Session 1 at 2026-10-20T09:11:49.6111191Z, private profile verified; QR claim and local runtime identity checks succeeded; curated visible record

The CLI bundle and native x64 payload above were previously installed into E's owned standard-user profile over SSH. No daemon was started in that historical setup: the installer creates a desktop sign-in task, and Task Scheduler does not start the daemon over SSH. The combined desktop build is now complete with DRY_RUN=1: GCP authentication was skipped and there were no bucket reads or writes. This artifact-build result does not establish app or device acceptance.

C's earlier staging desktop app and completed owner-account setup are historical context. They do not prove E4.3's final combined dialog, lease or cancellation behavior. Use the combined desktop artifact above for E's run and retain its exact source identity. A read-only comparison of latest C 38dde against the tested C parent 87779178 found no delta in serial, flash, native or CLI-runner inputs; the later shortcut and clipboard changes were not merged here. This does not change the tested artifact provenance. The desktop zip and E's manifest-built CLI installation are separate inputs; the desktop artifact alone does not establish that E's CLI or serial helper is installed. Keep the tested development CLI from being replaced by an updater.

Evidence already available

The committed E1.3 serial measurement record is automated hardware evidence from an isolated x64 test-PC measurement. It is not a completed E5 scenario:

  • Asserted DTR/RTS attempts failed the banner predicate. The released-lines fallback's first open also contained ESP-ROM: and failed; preserve that result.
  • The next three consecutive opens on the unchanged fallback binary were banner-free. Each held the port for 10 seconds, exited 0 at stdin end and kept the USB attachment session unchanged. The measured binary came from source 86398c252ab16ad285120f9a0b0303c2922ff7a1, Actions run 39383869283, with SHA-256 3efe833879dcc25de881fd945a3c7b7152e29308eb88ec3c0c42992a0342c035.
  • The initial banner's origin and exact reset edge remain unproven. An unchanged USB session alone does not prove that the firmware did not reboot. Person screen observation and protocol/heartbeat continuity were NOT RUN in E1.3 and remain required in E5.
  • E1.3 preserved all five preexisting process identities, removed its scratch and left no E-owned processes. It installed no SSH key. These are historical cleanup results, not a substitute for a new E5 cleanup audit.

The local integration reports in the plan workspace record the following for the frozen integration source above. They are local test evidence, not native Windows or real-login/device acceptance:

Prerequisite check Recorded result Wall time
CLI typecheck PASS 20.266 s
Focused CLI cable/devices/scripts/architecture tests PASS; 975 passed, 40 skipped 178.528 s
Required CLI core/services coverage gate FAIL; 2,902 assertions passed, required per-file coverage failed in devices.ts 130.761 s
Desktop dependency resolution with enforced lock PASS 2.024 s
Selected seven-file desktop VM suite PASS; 100 passed, none skipped 61.300 s
Scoped desktop analyzer FAIL; 17 inherited diagnostics: 14 style infos and 3 unused-import warnings 15.740 s

The CLI validation window was 178.528 seconds because its checks ran together; do not add their durations. The desktop validator's dependency-start through VM-receipt interval was 86.398 seconds. Neither is total request-to-completion time. The integration review found no blocking issue in its scoped C/E merge contracts, and recorded an inherited summary-ordering follow-up. It did not review all incoming C code or exercise hardware.

The original failed coverage gate above is retained as history. On final source 4c3def3cccaf0e820de63aeeb742a9852c6ce95a, the corrected core/services gate PASS covered 198 files and 2,904 tests, with 100% statements, branches, functions and lines, in 111.447 seconds. Final TypeScript typecheck also PASS in 4.137 seconds. The reviewed inputs were unchanged during these checks. All scoped reviews passed; that does not expand their scope to hardware acceptance or erase the inherited analyzer diagnostics.

Old native run 39387636051 is terminal FAIL: both architectures failed the same PowerShell 5.1 fixture; the recorded result was 738 passed and 56 skipped, with installation skipped. The fixture correction is in ddee034120a436d742c4c5bd255ea9bdd7444989. Replacement native run 39390890060 targets the exact final source. Its x64 job 118521849703 succeeded: Windows CLI tests recorded 739 passed, 56 skipped (58 files passed, one skipped); flasher tests recorded 18/18 passed, and the real public-manifest --no-flash check for 0.0.92 passed. This is native x64 prerequisite evidence. The completed run now has all four jobs successful across x64 and ARM64, including actual watch startup and clean uninstall. These native checks do not establish real-dial flashing or complete E3 acceptance. The 40 Mac-skipped CLI cases are not Windows passes, and the failed analyzer remains visible as a validation limitation. No claim that all validation passes is made.

Setup and isolation recipe

This is a run recipe, not an execution transcript. Follow the design's testing contract and the E5.2 task brief, with the current ownership facts and build requirements in this report taking precedence over obsolete setup assumptions.

The historical E5 setup record captures the owned standard profile hxdial: CLI install exit 0, version/helper hash confirmed, no daemon started, and a console=false COM3 watch that exited 0 within 5 seconds after stdin closed. The post-install audit preserved all 5/5 preexisting PID/name/creation identities with zero additional processes. hxdial2 is also an E-owned standard profile, but has no CLI installation. Loopback assets passed hash, native PowerShell and Node syntax checks; the server was OFF and no hosts/certificate changes were made. Initial setup command timestamps were not captured; the record's 2026-10-20T08:19:12.446341+00:00 curation timestamp is not a run duration.

The owner has changed the visible E5 run to the owner account's existing active Windows desktop. Assemble a private portable E profile under E-owned temporary storage, using process-local USERPROFILE, APPDATA, LOCALAPPDATA and TEMP, with separate CLI/auth/data/computer identity. Give the app a unique HARNESS_DESKTOP_INSTANCE, set HARNESS_DESKTOP_INSTALL_ROOT to the owned app stage and launch its executable directly. Preserve the owner account's actual PATH, tasks, registry and installation state, including C's app. The private assembler passed native Windows PowerShell 5.1 parsing for all 11 files. The exact staged app executable was then verified visible in Session 1 with its private profile at 2026-10-20T09:11:49.6111191Z. This proves the isolated launch, not completion of the eleven scenarios.

Use the dedicated private Limited/Interactive task driver, which checks the intended session and default desktop and retains positive owned-root/task receipts outside the repository. Public kit new-user guards remain intact; they were neither weakened nor adopted for the owner account. Use only the role label "owner account" publicly: never retain its actual identifier, address, password or SID, including in ignored repository notes. Raw inventories, credentials, full device identities and full local paths also stay outside the repository.

Authentication completed through the normal new-Windows QR claim, approved using the existing Mac SSO bearer held only in memory through the supported backend. Mac authentication files were not copied. The target's claim succeeded; the expected account, its own authentication with method qr, and its separate computer identity were verified privately. A backend HTTP response or the initial Node login child's exit alone is not the claim proof. No credential, QR code or account identifier is retained here, and the Mac daemon was not mutated.

Before device operations, the authenticated local WebSocket probe verified the E runtime's computer identity, data/auth directories, core PID and daemon record, then targeted only its own machine. It found one attached USB dial on firmware 0.0.92, 14 reported setting names, and two owned serial helpers (watch plus open). E actions targeted only its owned runtime; no action against C was authorized or performed. The final shared-state comparison below is NOT PASS and does not certify unchanged C state.

Use the exact native/desktop artifacts above, verify their hashes and the portable runtime's identities, and capture reviewed redacted visible app evidence. Serve the branch flasher only from E-owned loopback assets when needed; no deployment is required. Run the planned order 1, 3, 2, 4, 5, 6, 7, 8, 9, 10, 11 and target only E-owned processes. Quit the owned app during the stop/hold-port scenarios so supervision cannot invalidate them. Fresh baseline and final audits must preserve preexisting work; never assume a zero-process machine.

The procedures below preserve the original acceptance requirements as reference. For the current automated run, record actual Script/UIA observations and mark physical Person-only screen, cable-pull or user-switch observations NOT RUN unless they really occur. Never infer a physical action from software state. Scenario 1 has attached-state prerequisite evidence; scenario 2's Script portion passed, and scenario 3 has partial evidence recorded below. Their full acceptance verdicts remain open. Scenario 5 passed its Script-only acceptance case; scenario 6's Script portion passed, with Person update-screen observation still NOT RUN. Scenario 7 has partial visible flash-success evidence, with return timing/session comparison open. Scenario 8's first cancellation was before observed byte writing, and the subsequent successful flash failed to restore attachment. A later genuine mid-write cancellation and direct Try again rerun also failed recovery on c95e86208. Round 2 has local code evidence, with native entry/upgrade and decisive normal-CLI recovery PASS. This combines the proven GUI cancellation with normal CLI recovery; a new dev3 GUI flash was not run. S11's standard-user SSH uninstall failed after partial cleanup. S4's PnP simulation is blocked and S10 plugged metrics have verified derived postprocessing; the remaining scenarios are NOT RUN.

Observation and timing rules

Label What it can establish
Script Timestamped command results, device protocol fields, logs, file checks and owned process identities. A script cannot establish what a person saw on the dial.
UIA An actual named desktop control action and its observed app state, supported by a reviewed screenshot. Record unavailable controls and a Person fallback explicitly.
Guarded visible automation Foreground/PID/session-checked coordinate input followed by a reviewed screenshot and independent state read. This is the current Devices/brightness evidence; it does not establish semantic UIA control coverage.
Person The owner physically observes the dial screen, pulls/replugs the cable or switches Windows users. Record the requested action, actual report and timestamp; do not infer a report from logs.

Use helper stamp timestamps for each action start/end, and capture the actual event used for a latency calculation (plug, port release, cancel or user switch). Report UTC wall-clock start/end and elapsed seconds, including time waiting for a Person separately. An unreceived Person observation stays NOT RUN.

For each scenario retain: final source/artifact identity, start/end, action time, observed completion time, elapsed time, Script/UIA/Person observations, redacted evidence paths, verdict, findings, and any owning fix commit plus rerun evidence. Evidence filenames below are planned names unless explicitly linked. Review and mask every screenshot before adding it to the repository. The current raw screenshots contain full device identity and private workspace information and remain private; no publishable image is available yet. The reviewed screenshot at 2026-10-20T09:38:12.0207855Z has SHA-256 f1575a6704cdfa9fcf61a3d1bede2280e58dbaa9ea48fd95dd26015d0b0a8e57. It was retrieved through the private allowlisted copy path and visually inspected; its hash identifies private evidence, not an attached public screenshot.

Acceptance summary

# Scenario and acceptance Required observation Result Timing / evidence
1 Plug into a running daemon: attached, fw, id within 5 s; dial shows this PC's agents Script + Person photo NOT RUN for timed physical hotplug; attached-state prerequisite observed One attached dial, fw=0.0.92; no hand-replug event, latency or Person photo; second dial NOT RUN
2 Opening does not reset; heartbeat uptime grows through two daemon restarts Script + Person screen observation Script PASS; Person screen observation NOT RUN; full scenario incomplete Uptime 8112 → 8172 → 8232, same dial attached after both actual restarts; 205.618 s, ended 2026-10-20T09:52:19.4693671Z; original private baseline failure retained
3 Devices tab Connected; settings and pet changes acknowledged and visible Script + visible automation + Person + screenshot PARTIAL; Connected and brightness verified; pet/physical screen NOT RUN Private reviewed UI shows 35% then 80%; fresh own-machine probe confirms 80 and attached; curated visible record
4 Device restart and physical replug change attachment session; reconnect within 5 s; hard reset keeps attachment session Script + Person BLOCKED: PnP simulation returned reboot-required 3010; not a product reconnect FAIL At 13:51:26.1423Z, monitor ADMIN_RESTART_NOT_PASSED, session unchanged; no reboot/retry, physical replug or hard-reset acceptance
5 Busy port is left alone with one warning; reconnect within 10 s of release Script PASS Separate owned .NET SerialPort held COM3 for 20 s; attached count 0 while held, exactly one warning; same dial attached 2.636 s after release; 32.725 s total
6 Older firmware flashes, newest is offered/transferred over cable, dial reconnects on newest Script + Person Script PASS; Person update-screen observation NOT RUN; full scenario incomplete CLI 0.0.91 flash exit 0; own-daemon 0.0.91 → updating 0.0.92 → 0.0.92 and complete protocol evidence; 111.333 s
7 App detects port and flashes newest; running <v> ✓; daemon/device return Script + visible automation/Person + screenshots PARTIAL; original GUI success and later fixed mixed-input flash observed; required CLI-return timing not established Original run: 0.0.92, app 1:01, later attached probe. Fixed mixed run: actual 27.5% writing, fresh hello/attached and unchanged watch session; native-writer-exit metric 21.3845 s, not the CLI-return acceptance metric
8 Cancel during write; daemon back within 10 s; subsequent flash succeeds Script + visible automation/Person + screenshots Scoped PASS: actual GUI mid-write Stop + normal CLI3 recovery; new dev3 GUI flash NOT RUN Writers gone in 1.317 s; daemon open activity 4.615 s (first IPC unmeasured); decisive CLI3 exit 0, negative verdict cleared, same dial reattached 2.7990 s after CLI exit
9 Fast user switching transfers the dial to the standard console user and back within 10 s Script + Person NOT RUN Owner active in visible Session 1 with C/shared tests; autonomous user switching was unsafe
10 Ten-minute idle CPU/RSS measurements with and without the dial, per daemon/helper Script + Person for cable action PARTIAL: plugged derived metrics verified; original collector postprocessing FAIL retained; unplugged NOT RUN Same 5 stable pairs; total CPU 20.953125 s including test loopback server, core 20.8125 s/serials 0.0625 s; snapshot wall 601.3578848 s; stopwatch-at-failure 601.7609357 s includes final probe/postprocessing
11 Owned uninstall leaves no helper processes/files; dial returns to waiting screen Script + Person FAIL for attempted standard-user SSH uninstall; cleanup partial Exit 1 after stopping own scheduled daemon/removing PATH; helper cleanup AccessDenied under standard SSH; later owned inventory: 0 processes, 1 serial file, launcher present; desktop completion/waiting screen not done

Scenario procedure and result fields

The procedures remain the acceptance recipe. Only the partial observations explicitly recorded below have run; unrecorded historical result/timing fields are labeled Not recorded.

1. Plug in with the daemon already running

Person unplugs the dial. Record the owned daemon's running state and no attached device, then Person plugs it in and supplies the event time. Poll device state and measure first attached with firmware and ID against the 5-second limit. Retain a Devices screenshot and a Person photo showing this PC's agents on the dial. Planned evidence: dial-s1-before.log, dial-s1-after.log, dial-s1-devices.png and a redacted dial photo. A second dial is a separate optional fleet check; record availability and outcome without treating one dial as multi-device evidence.

Current partial observation: the owned runtime reports one attached dial with firmware 0.0.92, and the reviewed Devices UI shows one Connected card. No physical unplug/replug, arrival timestamp, five-second measurement or dial-screen photo has been recorded. Scenario 1 and the second-dial check remain NOT RUN.

2. Opening does not reset

Carry E1.3's failed first fallback open and three subsequent banner-free opens with their exact source limits. Obtain the missing Person screen observation. With the owned app quit and the dial attached, capture alive up=<s> heartbeat lines before and after each of two owned-daemon stop/start cycles. Allow enough time for the minute heartbeat and record reattachment each time. Uptime must increase across both restarts; unchanged USB session alone is insufficient. Planned evidence: dial-s2.log, redacted heartbeat excerpts and a timestamped Person report.

Current partial observation: before any restart test, firmware heartbeat uptimes were 6244, 6304, 6364, 6424, 6484, 6545, 6605 seconds, strictly increasing. The E app was quit through its own Harness menu at 2026-10-20T09:43:25.4972147Z; only that app was verified absent and the owned daemon still reported the dial attached. The first private restart-test attempt failed its baseline predicate after 241 ms, before any stop/start: PowerShell 5.1 returned a singleton device object whose scalar .Count was null. Native inspection confirmed one attached device with an ID; array normalization fixed only the private test predicate. The original failed receipt was preserved. The corrected native private test then completed two actual owned-daemon restarts: baseline uptime 8112, first-restart uptime 8172, second-restart uptime 8232 seconds. The same owned USB dial was attached (count 1) after each restart, and the own endpoint identity was verified before each stop/read. Four checkpoints completed, with recovery-daemon verification true. The full private JSON receipt was downloaded and read by the coordinator. The test finished at 2026-10-20T09:52:19.4693671Z in 205.618 seconds.

This is Script PASS for two-restart continuity and recovery. The earlier pre-test heartbeats alone did not establish that result. Person observation of the physical dial screen is still NOT RUN, so the full scenario remains incomplete.

3. Devices tab

Open Devices using UIA where named controls exist, otherwise Person. Check Connected, change a setting reported by the dial and apply a pet. Capture the returned settings/device state and a screenshot with Windows-appropriate copy. Person confirms the setting and pet on the dial itself. Planned evidence: dial-s3.log, dial-s3-devices.png and the Person observation.

Current partial observation: a reviewed private Devices screenshot shows one Connected card. The detail screenshot at 2026-10-20T09:35:03.3470782Z shows Connected, firmware 0.0.92 and brightness 35%. A protocol operation changed brightness 80 → 35. A guarded foreground/PID/session coordinate interaction then moved the actual UI slider 35 → 80; a fresh own-machine firmware probe confirmed brightness 80 and attached=true. The visually inspected private screenshot at 2026-10-20T09:38:12.0207855Z shows 80%. There were 14 reported setting names. No pet change or Person confirmation on the dial has run.

UIA exposed only two root nodes, so these actions are recorded as guarded visible automation, not named-control UIA coverage. A private legacy tool observer timed out because broadcasts skip tools; a fresh reconnect read of 35 confirmed the native setting write. The private observer was changed to use a correlated own-machine RPC. This was an observation-harness issue, not a recorded product failure.

4. Unplug, replug and reset

Privately identify only the test dial's composite USB device. Capture a redacted attachment session before and after a targeted device restart; never issue a broad device restart. Measure reconnect against 5 seconds and confirm the owned daemon stays responsive. Repeat for a Person's physical unplug/replug. Verify that an esptool hard reset/reflash retains the attachment session during scenario 7 and cross-reference that evidence here. Planned evidence: dial-s4.log, event timestamps and scenario 7's session comparison. Keep full instance IDs outside repository output.

Current result: BLOCKED. A fresh CLI3 monitor ran, but the targeted PnP simulation returned PNPUTIL_REBOOT_REQUIRED_3010 / exit 3010 at 2026-10-20T13:51:26.1423Z. The monitor reported ADMIN_RESTART_NOT_PASSED and attachment session unchanged. No reboot or retry was performed. Physical hand replug and hard-reset acceptance did not run; this blocked script cannot establish a product serial-reconnect failure.

5. Port held by another program

With the owned app quit, stop only its daemon and use the owned port holder to hold the confirmed dial port. Start the owned daemon. Verify exactly one busy warning and no attached connection while held. Record actual release time and reattachment within 10 seconds. Planned evidence: dial-s5-held.log, dial-s5-free.log and redacted warning/timing excerpts. Clean up only the owned holder, even if the test fails.

Result: PASS. A separate owned .NET SerialPort held COM3 for 20 seconds. While held, attached count was 0 and the owned daemon identity was verified; exactly one busy report was emitted. After release, the same owned physical dial reattached (count 1) within 2,636 ms, below the 10-second requirement. The busy-report count stayed at one, and recovery-daemon verification was true. All 12 command exits were zero. The coordinator downloaded and read the private JSON receipt. Start: 2026-10-20T09:56:35.3556648Z; end: 2026-10-20T09:57:08.0729145Z; recorded duration: 32.725 seconds. This scenario has no additional Person predicate.

6. Firmware over the cable

Record the public manifest's newest firmware and a verified older release. With the owned app quit, flash that older release through the owned CLI, then observe fw=<older>, the newest cable update and fw=<newest> with attachment restored. Retain cable transfer logs and Person's update observation. Planned evidence: dial-s6.log and a redacted firmware/version record. A flash failure remains a failure until the owning fix and a successful rerun are recorded.

Current preparation: a read-only official firmware listing returned 92 versions; 0.0.91 was selected as the previous release. Live official metadata retrieved through HTTPS identifies latest 0.0.92, size 3,360,672 bytes, SHA-256 eadb3846a6658026048418804a60341adc45fdb24793ebc9cabd782def0c8458. The optional no-flash prefetch ended with a private PowerShell 5.1 NativeCommandError while handling expected older-version stderr. No board write occurred and no cache images were found; no prefetch rerun is planned. This preparation-harness failure is separate from a product flash outcome. The private S6 runner's per-dial proof matching was corrected before the actual run below. The earlier prefetch failure does not establish a product failure.

Result: Script PASS on source 4c3def3c. The owned CLI flash --version 0.0.91 --yes exited 0 in 85.468 seconds, reported the older version running and emitted the expected unverified-old-version warning. Fresh own-daemon samples recorded:

UTC Attached Firmware Updating
2026-10-20T10:13:49.4747816Z true 0.0.91 —
2026-10-20T10:13:58.9149062Z true 0.0.91 0.0.92
2026-10-20T10:14:11.0768362Z true 0.0.92 —

Fresh per-dial protocol evidence contained the old-version hello, the 0.0.92 offer (3,360,672 bytes), the complete installed acknowledgement (3,360,672/3,360,672 bytes) and the new 0.0.92 hello. Protocol completeness and observed-transition predicates passed. All probe commands exited 0; the owned recovery daemon was verified and no manual recovery was attempted. Polling took 24.4006 seconds. Start: 2026-10-20T10:12:20.4135623Z; end: 2026-10-20T10:14:11.7465309Z; elapsed approximately 111.333 seconds. Only allowlisted result fields were read from the private JSON receipt; raw protocol lines contain full device identity and are not copied here.

Person observation of the physical update screen is NOT RUN, so the full scenario remains incomplete. Scenario 7's separate partial GUI result is below; scenario 8's cancellation observation and subsequent recovery failure are below.

7. Flash from the app

Start the combined owned app. Record attachment session, open Flash firmware and capture detection, active writing and completion. Check that detection finds the confirmed dial port, the flash identifies ESP32-S3, writes newest and reports running <v> ✓. Check daemon/device recovery within 15 seconds and the wording "this computer". Compare attachment session before/after the hard reset and reflash. Planned evidence: dial-s7.log, dial-s7-probe.png, dial-s7-flashing.png, dial-s7-done.png and a Person report.

Current result: PARTIAL. The exact staged app on source 4c3def3c was re-launched visibly in its private profile, PID 2000 / Session 1, at 2026-10-20T10:15:31.1528052Z. Before the flash click, its actual native GUI probe showed COM3 ready. The private screenshot at 2026-10-20T10:20:37.0310449Z shows Flasher targeting 0.0.92 → COM3, with actual write progress at 21%. The coordinator reviewed the completion screenshot: title "circle 0.0.92 is on the dial", board COM3, flashed version 0.0.92, and "Took 1:01". Its output contains running 0.0.92 ✓ and starting Harness again. Screenshots and their metadata remain private.

This first attempt did not measure the <=15-second daemon-return bound or capture attachment-session identity before/after. A fresh own-runtime read-only probe started at 2026-10-20T10:24:15.041Z returned code OK, verified the owned daemon and found one attached dial on firmware 0.0.92. This later recovery observation does not measure the <=15-second return bound. The visible success and the app's duration display alone do not complete scenario 7 or establish those missing predicates. Physical dial observation is also NOT RUN.

Fixed mixed-input observation: after the verified private upgrade, a new actual GUI flash ran with app/native source 4c3def3c and CLI/PowerShell source c95e86208. A reviewed private image showed 27.5% actual writing. Its read-only observer passed, observed a fresh protocol hello and attachment, and reported unchanged before/after watch attachment session. The native writer exited at 2026-10-20T12:38:32.3713346Z; the completed fresh query observed attachment at 2026-10-20T12:38:53.7558803Z, a difference of 21,384.5457 ms. The observer ran from 2026-10-20T12:37:37.5504048Z to 2026-10-20T12:38:54.0641423Z. This metric starts at the OS native-writer exit, not CLI completion: it establishes neither PASS nor FAIL for the required 15-second CLI-return criterion. The completion image at 2026-10-20T12:41:32.0356658Z was reviewed (SHA-256 4cd0c3c718e1234b51967d7cee4df2eab537a60f112433c0bf227dbdc7146318); the own query at 2026-10-20T12:41:13.843Z returned OK, signed in, daemon verified, dev.dial2 and one attached dial.

The fixed-run baseline was already attached, with a different core PID from the original failure; why the old unattached state ceased is unproven. This flash does not establish the decisive scenario 8 regression result. The owner changed the network endpoint, with SSH host-key continuity verified for the same test PC. The first window-preparation timeout after that transition was preserved; second preparation and fresh screenshot passed, requiring no UIA fix. No endpoint address, host-key fingerprint or raw private image is included.

Historical M1 observation: this app artifact displayed flash-circle.sh while running PowerShell flash-circle.ps1. Source fix 444139f69 later passed widget checks and scoped rereview; that does not update the earlier GUI artifact/image.

8. Cancel a flash

Start another app flash and cancel during write app, recording the cancel event time. Check cancellation within 2 seconds, daemon back within 10 seconds and, at 15 seconds, no owned esptool process and only expected owned daemon serial helpers. The binding spec requires the daemon back; the plan's extra immediate attached=true requirement conflicts with its allowance for a torn firmware image and does not govern this cancellation step. Record daemon-log activity separately from the first successful IPC response. A cancelled write may leave no runnable image; perform a second app flash and verify success on newest firmware with Person confirmation. Planned evidence: dial-s8-cancelled.png, recovery timestamps, owned process audit and second-flash evidence. No cancellation success is inferred from a unit-test result.

First cancellation observation: PARTIAL, not a confirmed mid-byte-write cancellation. At 2026-10-20T10:27:40.9299071Z, the guarded click targeted Stop while two exact owned COM3 write_flash processes existed. Both were gone at 2026-10-20T10:27:41.4869419Z, with recorded disappearance latency 559 ms; zero owned esptool processes remained at 2026-10-20T10:27:56.1025101Z. The private watcher completed without an error. Its process checks do not by themselves establish when the UI cancellation completed or the daemon returned.

The coordinator-reviewed private screenshot shows "Flash failed", "Stopped. Harness has been started again." and a cancellation error in the output. It also shows esptool at Connecting...: byte writing was not observed before the click. This attempt therefore does not satisfy the during-write predicate. The subsequent Try again recovery flash completed successfully in the actual GUI: running 0.0.92 and "Took 0:59". Recovery result: FAIL. Fresh identity-verified, read-only queries at 10:31 and 10:34 UTC still reported attached count 0 and firmware null. Successful flashing did not restore the device connection.

At 10:37:34 UTC, private diagnostics found no active flashing or cancellation marker, only the E-owned serial watch process (no open helper or esptool). The owned per-dial log opened at 2026-10-20T10:27:45.425Z and closed for silence at 2026-10-20T10:28:05.520Z. E's persisted data/dial-ports.json held one foreign-port verdict dated 2026-10-20T10:28:05.505Z, tied to the unchanged USB attachment session. No new hello had appeared since 2026-10-20T10:21:21.245Z, and no busy-to-free recovery line was present. No full device identity or raw protocol line is retained here.

Source diagnosis confirmed that the new daemon reloads this six-hour foreign verdict and skips the port. It was off during the successful reflash and missed the busy-to-free transition; the unchanged USB attachment did not invalidate the verdict. This is a product recovery failure, separate from the earlier private observation/preparation-harness failures.

The first scoped CLI completion-hook code change was implemented at c95e86208dafb484ee9a662972b7a196bffaa6d3: after a successful write and running-version verification, before PowerShell restarts the daemon, it was intended to clear only the chosen port's verdict while preserving siblings. Cancelled, failed, no-flash and unverified-running paths retain their verdicts. The task's scoped review is clean. Local CLI checks passed 73/73 plus TypeScript typecheck; actual native test-PC PowerShell 5.1 fixtures passed 13/13 with no hardware or network operations. These checks establish code/fixture behavior, not recovery of the original failure. Native workflow 39412223533 now has all four jobs SUCCESS: each architecture passed 739 contract tests (56 skipped), 31/31 flasher checks plus official 0.0.92 no-flash, exact watch installation and clean uninstall.

The private CLI upgrade passed native configuration/current-version/helper-pin checks, preserving the before/after verdict file hash. App/native artifacts remain 4c3def3c; CLI/PowerShell are c95e86208. The later fixed-build flash observer is documented under scenario 7. Its baseline already showed the same dial attached and a different core PID from the original failure; the earlier unattached state had ceased before that write, for an unproven reason. The unchanged verdict hash and successful ordinary flash therefore do not prove that the S8 negative-verdict recovery failure is fixed in hardware.

The prior 4c3def3c failure remains recorded (failure-receipt local head 030ba49cf). Earlier S2, S5 and S6 receipts retain their original source scope. An intermediate private v2 watcher missed the write window, performed no Stop and rejected a stale arm request; that is a preserved harness attempt, not a product failure.

The later timed cancellation is genuine: the reviewed pre-click image at 2026-10-20T13:01:49.3269075Z shows Writing 17.0%, 212,992/1,249,692 compressed bytes. Stop occurred at 2026-10-20T13:01:50.3652759Z with two exact owned writers. Both were gone at 2026-10-20T13:01:51.6806950Z, recorded latency 1,317 ms (within 2 seconds); zero remained at 2026-10-20T13:02:05.4872051Z. The watcher completed without error. A reviewed post-cancel UI image at 13:04:10.9447 UTC shows Stopped/Harness started again, a cancellation error, and last Writing progress 19.7%.

The owned per-dial log opened at 2026-10-20T13:01:54.980Z, 4.615 seconds after Stop, then closed for silence at 2026-10-20T13:02:15.066Z. This proves daemon activity at 4.615 seconds, not the time of its first IPC response. Post-cancel unattachment is expected for a torn image and is not itself the recovery failure. The negative-verdict receipt records one verdict at 2026-10-20T13:02:15.055Z, the same USB attachment session and false markers.

The decisive direct Try again recovery used that unattached baseline and made no new detection, cache clear or PnP intervention. It started at 2026-10-20T13:08:18.839Z; the native writer exited at 2026-10-20T13:09:12.6776Z. At 13:10:36.69 UTC, the old negative verdict still persisted, with no new open or hello. The actual completion UI reviewed at 13:12 UTC reports running 0.0.92 ✓, "Took 0:59" and no callback warning, but the dial remains unattached. The c95e86208 hardware recovery result is FAIL. The prior ordinary fixed-build flash, whose baseline was already attached, remains noncausal evidence.

Round 2 is committed at 4fe12de12d7e1d8695bae0b01529a432fec28a17 and passed independent scoped review with no new breakage found. The lifetime regression using the actual watch and short-lived Node process recorded RED: 3 failed, 9 passed; the final selected suite recorded GREEN: 103/103, with TypeScript typecheck PASS. Actual Windows PowerShell 5.1 bundle-entry fixture V2 passed: the old dev2 exited 0 with both verdicts unchanged and no fake start (not a full watch RED); CLI3's 500 ms snapshot showed only selected A removed, B unchanged, exit 0 and its helper gone within 5 seconds. CLI3 selection upgrade also passed, with verdict bytes and the same 13:02 foreign fixture unchanged. Its bundle checksum is recorded above. The subsequent decisive normal-CLI hardware recovery passed as recorded below. Historical 13/13 fixtures and green CI remain scoped to the first revision and do not erase the hardware failure. A private orchestration v1 failed after starting only the read-only observer, on a PowerShell 5.1 File.Replace null-backup issue. A read-only partial receipt confirmed no flash request, zero writers, false markers, an unattached baseline and the same single negative verdict. The private helper v2 correction was still being prepared at that point; this superseded attempt is not a product failure and performed no CLI3 hardware flash. The later normal CLI recovery below did run; the later blocked S4 and verified derived S10 results are recorded in their own sections.

Decisive CLI3 recovery result: PASS on app/native 4c3def3c, CLI 4fe12de12 / 0.9.0-dev.dial3 (bundle 8cacd1f7…) and PowerShell c95e86208. The observer baseline was unattached with the unchanged 13:02 negative verdict. There was no preparatory detection, cache clear, PnP operation or GUI action. The normal product command flash --yes --port COM3 ran in Limited Session 1 while the existing app remained on Done, avoiding a new GUI probe.

The product CLI started at 2026-10-20T13:42:28.9149Z, exited at 2026-10-20T13:43:31.8549Z with exit 0 in 62.960 seconds, and reported running 0.0.92 with no completion warning. Driver time was 64.775 seconds. The selected negative verdict count became 0, both markers were false, the same app stayed alive, and a fresh hello/attachment was observed with the same USB attachment session. Native writer exit was 2026-10-20T13:43:14.0166159Z; attachment was observed at 2026-10-20T13:43:34.6539207Z. That is 20.6373 seconds from native writer exit, and 2.7990 seconds from product CLI exit, the comparable 15-second return metric. The observer ended at 2026-10-20T13:43:34.9449923Z with PASS.

This is causal hardware recovery evidence for round 2 from the preserved negative state; the earlier already-attached ordinary flash was not. S8's recorded scope combines the prior real GUI cancellation at Writing 17% with this normal-CLI recovery. A new dev3 GUI flash was not performed and no such image is claimed. First IPC response after cancellation remains unmeasured; the recorded 4.615-second open proves daemon activity only. Earlier failed attempts stay visible above.

9. Two users

Create a second owned standard profile only if needed and record that ownership privately. Person uses fast user switching to put it at the console; install and start only its E test runtime. Capture both owned users' states: the active console user attaches and the other releases with the expected ownership log. Person switches back and the direction reverses. Measure each transfer against 10 seconds. Planned evidence: dial-s9-second.log, dial-s9-back.log and timestamped Person reports. The existing setup performed no desktop switch and cannot satisfy this scenario. It was not run because the owner's visible Session 1 also hosts C/shared tests, making an autonomous user switch unsafe. Do not switch, stop or alter a preexisting user.

10. Idle cost

With the owned app quit, record daemon and helper identities, CPU seconds and working set/RSS at the start and end of a 600-second interval with the dial. Repeat without the dial after Person confirms unplugging. Attribute each sample to the same owned process identity; report restarts rather than subtracting counters from different processes. Report CPU delta and memory samples per process for both intervals; there is no acceptance threshold. Investigate an unexpected number before declaring the measurement complete. Planned evidence: dial-s10.log and the table below.

Current status: the app was quit normally and the private E backend remained. The plugged snapshots captured the same five stable process pairs at 2026-10-20T13:53:12.1757386Z and 2026-10-20T14:03:13.5336234Z, a wall interval of 601.3578848 seconds. The original collector FAILED postprocessing on Measure-Object comparison of an OrderedDictionary. That receipt is preserved unchanged, SHA-256 f7cd287039026718a92275840d1fa2f3bee73cc59c9241ec2669be2b378f62b1; it is not relabeled PASS.

Separate strict native postprocessing of those same captured vectors PASS: the derived receipt reports verified=true, postprocessing_only=true and unchanged source data. Verified total CPU delta is 20.953125 seconds, including the test loopback server; core delta is 20.8125 seconds and the two serial helpers total 0.0625 seconds. The pinned collector initially assigns elapsed before its final probe, but its catch block overwrites elapsed after the failed comparison. Thus 601.7609357 seconds is stopwatch elapsed at failure, including the final probe and failed postprocessing, not the clean snapshot interval. The final probe ran 14:03:13.5616602Z → 14:03:13.7529415Z; the collector ended at 14:03:13.9389074Z. The derived receipt explicitly sets elapsed_at_failure_includes_final_probe_and_postprocessing=true. No clock-adjustment explanation is inferred. No 600-second rerun occurred; there is no performance threshold and no unplugged sample.

Dial state Process role Start UTC End UTC Snapshot wall s Verified CPU delta s RSS start/end bytes Finding
Connected Core 13:53:12.1757386Z 14:03:13.5336234Z 601.3578848 20.8125 93790208 → 78700544 Verified derived, same stable pair
Connected Master 13:53:12.1757386Z 14:03:13.5336234Z 601.3578848 0.078125 27111424 → 15917056 Verified derived, same stable pair
Connected Serial helper A 13:53:12.1757386Z 14:03:13.5336234Z 601.3578848 0.046875 5238784 → 5165056 Verified derived, same stable pair
Connected Serial helper B 13:53:12.1757386Z 14:03:13.5336234Z 601.3578848 0.015625 6103040 → 5918720 Verified derived, same stable pair
Connected Test loopback server 13:53:12.1757386Z 14:03:13.5336234Z 601.3578848 0 13942784 → 14012416 Included in total 20.953125 s
Disconnected No sample Not recorded Not recorded Not recorded Not recorded Not recorded NOT RUN

11. Uninstall

Run uninstall only in each E-owned test profile. Verify successful exit, zero owned harness-serial processes and no helper files in that owned installation. Person observes the dial's waiting screen. Planned evidence: dial-s11.log, owned file/process audit and a Person report. Preserve preexisting Harness installations, including C's installation; global process-name checks cannot justify removing them.

Current result: FAIL for the standard-user SSH attempt; cleanup partial. The actual uninstall exited 1. It stopped its own scheduled daemon and removed its PATH entry, then refused helper cleanup with AccessDenied under the documented standard-SSH restriction. A private diagnostic directs uninstall from the Windows desktop. No uninstall replay has been attempted. A later administrator read-only ownership and inventory audit completed: that installed test profile has zero owned processes, one serial-helper file and a launcher still present, confirming incomplete removal. The second, tools-only profile has zero owned processes, zero serial files and no launcher: its read-only inventory passed, but no uninstall ran there because no product was installed. Both profile-removal attempts later stopped at the Loaded guard with no account/profile deletion; ownership receipts and shared scratch remain retained under the owner cleanup waiver. The physical dial waiting-screen observation and owner-E product uninstall are NOT RUN; its separate process/task teardown is recorded below.

An earlier private invocation was blocked by PowerShell ExecutionPolicy before any uninstall action. Process-only Bypass corrected that harness invocation; it is separate from the subsequent actual uninstall failure above.

Findings, fixes and limits

Item Status / required follow-up
Core per-file coverage gate Initial failure retained; corrected final gate PASS, 198 files / 2,904 tests, all four coverage metrics 100%, 111.447 s
Final TypeScript typecheck PASS on unchanged reviewed final inputs, 4.137 s
Earlier scoped reviews PASS within their original scopes; whole final review returned with fixes below
Native PowerShell fixture / final native CI Old run 39387636051 terminal FAIL; fixture corrected in ddee0341; final run 39390890060 all four jobs SUCCESS across x64/ARM64
Final CLI bundle and installation 4c3def3c bundle PASS in 15.7 s; owned SSH install exit 0/version/helper hash confirmed; no daemon started
Combined desktop artifact Run 39393250390 both architectures SUCCESS, dry run only; isolated visible app and QR claim verified; eleven-scenario acceptance incomplete
E5 setup watch console=false, COM3; stdin-close exit 0 within 5 s; no COM open/probe and no acceptance conclusion
Historical loopback asset setup At initial setup, hash/native parse/Node checks passed with server OFF and no hosts/certificate changes; later test server is included in idle inventory and its stop passed; only profile/file cleanup was blocked and is now owner-waived
Desktop analyzer 17 inherited diagnostics recorded; analyzer failed
Final review I2 Inherited Important summary ordering in cli/src/services/devices.ts:309-319, outside E/nonblocking for E; remains open for notification/C owner and is carried to the owner final report
E1.3 first released-lines banner Failed measurement retained; reset origin unproven
Person screen/heartbeat proof Two actual restarts: Script PASS, uptime 8112 → 8172 → 8232, attached after each; Person physical screen observation NOT RUN
Private observation harness Legacy tool observer timeout and singleton .Count baseline failure preserved; corrected privately, neither establishes a product failure
Firmware scenario 6 Script PASS: actual CLI older 0.0.91 flash, observed cable transition to 0.0.92, complete offer/installed/new-hello proof, owned daemon recovered without manual intervention; Person screen observation NOT RUN
Historical optional firmware prefetch Private PowerShell stderr handling error; prefetch made no board write and found no cache images; no prefetch rerun planned; separate from the successful S6 run
Devices UI and settings Connected and brightness 35 → 80 verified in actual UI and fresh own-machine state; pet and physical confirmation NOT RUN
Devices chat panel Shows not ready because backend chat could not start; observed separate issue, root cause unresolved
Busy-port scenario 5 PASS; held 20 s, exactly one warning, attached 0 while held, same dial reattached 2.636 s after release; recovery verified; 12 command exits zero
App flash scenario 7 Original GUI success remains partial. Fixed mixed-input run: actual writing, fresh hello/attachment and unchanged watch session; 21.3845 s measured from native writer exit is a different metric from required CLI-return timing; reviewed completion image and own query confirm signed-in dev.dial2/verified daemon/attached 1
Cancellation/recovery scenario 8 Scoped PASS: real GUI cancellation + decisive normal CLI3 recovery from preserved negative state, cleared verdict and same-session fresh hello/attachment 2.7990 s after CLI exit. Dev3 GUI flash NOT RUN; earlier 4c3def3c/c95e86208 failures retained
First recovery change / round 2 c95e86208 native fixtures/CI passed but hardware recovery FAILED. Round 2 4fe12de12: actual-watch/short-lived Node RED 3 failed/9 passed; final suite 103/103 + typecheck PASS; independent review PASS; native entry/upgrade and decisive normal-CLI hardware recovery PASS; dev3 GUI flash NOT RUN
Standard-user SSH uninstall / inventories Actual exit 1; later complete read-only audit: installed profile has 0 owned processes, 1 serial file and launcher present. Tools-only profile inventory PASS (0/0/no launcher), with no uninstall. Both profiles retained; owner-E phase removed no paths or installation
S4 PnP simulation Blocked by native reboot-required exit 3010; monitor did not pass; no reboot/retry or physical acceptance; no product reconnect failure inferred
Whole-implementation final review / I1 Original review at 4fe12de12 returned with Important cancellation/restart blocker. I1/M1 committed in eight paths at 444139f69; scoped rereview PASS resolved I1/M1 with no new findings. Native PS5.1 function-only fixtures 7/7 PASS; no new GUI hardware result
Final-wave focused validation 84 passed/1 existing skip assembled from two receipts: final flasher 23/23 plus five completed files 61 passed/1 skip; not an uninterrupted six-file passing run. Scoped analyzer, CLI wrapper and website checks passed within report scopes
Final-wave compatibility limit New app with an older cached PS1 conservatively retains that script's legacy shared marker after forced retirement; the old 10-minute suppression can remain. No mixed-version forced-recovery claim
Final review M2 serial/tests/windows.rs:24-25 unbounded initial read remains deferred to combined Windows owner by recorded ruling; passing native runs do not resolve this test-robustness follow-up
Final review M1: Windows flasher header Source now uses the platform script name at 444139f69; widget checks and scoped rereview PASS; I1/M1 resolved with no new findings. The actual app artifact remains 4c3def3c, with no new GUI validation
E5 scenario outcomes and owning fixes Scenario 5 PASS; S2/S6 Script PASS; S7 partial; S8 scoped GUI-cancel/normal-CLI3 recovery PASS, earlier failures retained; S11 SSH uninstall FAIL/partial; header source fix reviewed; other full acceptance open
Second-dial fleet check NOT RUN; only one attached dial observed
Arm64 real hardware Not covered by this x64 hardware report

Any discovered fix belongs to its owning task/file, with a meaningful regression check and scenario rerun. Record the original failure as well as the replacement evidence; do not overwrite a failed run with a passing label.

Cleanup and final safety audit

Cleanup outcome: WAIVED BY OWNER on 2026-10-20. The failed account/path cleanup is accepted as a recorded limit; no further cleanup, reboot, hive unload or deletion attempts will run. Retain both test profiles, test keys, portable E root, shared scratch and positive ownership receipts. The cost is retained test accounts/files and disk space, with possible later manual cleanup. Historical outcomes follow. The owner-E v4 process/task teardown PASS: normal private CLI3 stop, exact held PTY and server stops, and 24 completed owned NoWait tasks unregistered. Final owned-root process and task counts were 0. This ran 2026-10-20T14:35:20.8387298Z → 2026-10-20T14:37:27.2000336Z, with receipt elapsed 126.472 seconds; receipt SHA-256 0a3b13b17ea4bbfcd57ae33e784746e1bea597bdfc7f6b98bd345ecb7f53ed68. No path deletion, product uninstall or owner-profile change occurred in that phase. Earlier cleanup guard/preflight failure, including v3, performed zero actions and is preserved separately. Both tools-only and installed-profile cleanup attempts ended BLOCKED by the Loaded guard. Each found 0 owned processes and performed no account/profile deletion; ownership receipts remain. Exact-profile read-only audits passed: each took 31 snapshots (20.988 and 20.919 seconds), verified exact owned identity/path, found Loaded=true and Special=false with both hives present, and could not determine RefCount. No exact-identity logon or WTS session was present, so no safe owned session logoff was available. No hive unmount has been performed. Removal attempts were BLOCKED; the owner has waived further cleanup. Loader/handle provenance is unproved; no direct hive unload, global service restart or reboot was performed in this run. The prepared remover was not executed. Shared scratch remains retained under the owner waiver.

The five-proof native ZIP was built after the owner-E process stop and before any path deletion. The preexisting private Mac proof manifest preceded the stop. Native archive and download checks passed; ZIP SHA-256 b9dd06eecfe2348b65076dd3e998421c0657f0775906f5937e6acddc11963fe9, 11,830 bytes, five source files. On the Mac, whole-ZIP hash, all five entry hashes and CRC were verified. This archive contains no authentication files, screenshots, configuration, profiles or raw command logs. Existing private scenario JSON/images/command proofs were already manifested before teardown; none of that raw content is copied here. The fresh read-only public-preflight comparison against the original baseline is NOT PASS: baseline 5 process identities/2 tasks; 3 old process identities are absent and 3 old listeners are absent (6 differences), while task differences are 0 and both tasks remain. Cause is unproven; C/shared work occurred concurrently. No E action against C was authorized or performed, and no baseline rewrite or C restoration was attempted. This does not certify full C state or establish that E caused the changes. It is separate from the guarded owner-E process/task teardown PASS.

Native inventory V2 completed with 12 owned processes, 109 task receipts and 0 inventory issues; this is inventory, not cleanup. The first inventory hit its 55-second deadline before enumeration, so its zero process count was not a measured zero. Owned paths remain retained; native function fixtures and owner-E process/task teardown are complete, remaining account/path cleanup was blocked by Loaded profiles and is now owner-waived. Preserve the private ownership ledger until cleanup is verified. Stop/remove only E-owned processes, loopback server, scheduled tasks, temporary files, installations, test users and newly created profiles. Remove only the dedicated test key from the owned profile and remove its private local copy through the private workflow. No administrator key was installed, so the historical task-brief command editing the administrator authorized-keys file does not apply. Never edit a preexisting key file on that assumption.

Compare the final private audit with the fresh E5 baseline. Confirm every preexisting process identity and C installation is preserved, all owned scratch, tasks/users/profiles/keys are removed, and no owned processes remain. Do not delete an object merely because its name matches a test name; verify ownership first. Preserve reviewed redacted evidence only. Record actual cleanup findings and any remaining owned objects rather than reporting a clean machine from a zero-count assumption.

Cleanup check Result
Original baseline comparison NOT PASS: 3/5 old process identities and 3 old listeners absent; 6 differences. Task diff 0, original 2/2 tasks retained. Historical initial setup 5/5 receipt remains separate; cause unproven
C/shared-state preservation Not certified by final comparison; concurrent C/shared work occurred, no E action against C authorized/performed, and no causal attribution or restoration
E-owned processes / loopback server / tasks removed Owner-E v4 PASS: normal CLI3/exact PTY/server stop; 24 completed owned tasks removed; final owned-root processes/tasks 0. This is not a global system zero or uninstall result
E-owned files / test installations / scratch retained Standard-user uninstall removed PATH entry then helper cleanup AccessDenied/exit 1; later audit confirms 1 serial file and launcher remain there; tools-only profile has 0 serial files/no launcher; owner-E phase removed no paths or installation
E-owned users and newly created profiles removed Both attempts historically BLOCKED by Loaded guard; owner waived further cleanup, each with 0 owned processes; no account/profile deletion; ownership receipts retained; exact-profile audits PASS but Loaded=true/no owned logon or WTS session, so no safe session logoff
Dedicated test keys Retained with ownership receipts under owner waiver; no removal claimed
Reviewed, redacted cleanup evidence Owner-E teardown receipt PASS retained; both Loaded-blocked account receipts retained; final baseline comparison NOT PASS; remaining cleanup open