Skip to content

armgit / kiln

Public Imported from GitHub
0

Build and revise procedural 3D assets with your coding agent. Local MCP server, CLI and TypeScript engine with rendered review, editable source and GLB export.

docs: reconcile shipped status and retire Troy review workflow (#143)

Matthew Kissinger42fab36d700e
diff --git a/ROADMAP.md b/ROADMAP.md
index a20b501c..699fb815 100644
--- a/ROADMAP.md
+++ b/ROADMAP.md
@@ -1,168 +1,106 @@
# Kiln roadmap
-**Release direction, 3 October:** Matt confirmed the farmer hand fixes and could not
-reproduce the water jump locally. He authorized final performance review, commits,
-main integration and Cloudflare deployment. The focused performance investigation is
-settled for release with its original failures and limits preserved. See the
-[release disposition](docs/reviews/2026-10-03-release-disposition.md). Bridge stills are refreshed; physical-device
-promotion, Troy construction and npm publication remain deferred.
-
-Kiln **0.10.0** is the current release, the v1 readiness release: the stdio server serves
-protocol revision 2026-07-28 beside the 2025 handshakes and answers before the engine
-loads, the three record tools are flat objects with an `action` field, every tool result is
-bounded and leads with its verdict, and the generated workspace guide and skills are
-shorter. It builds on 0.9.0's standalone authoring with optional projects, a material
-library and Live Review in the packaged local dashboard and the calibrated
-`review-neutral-v1` review rig. [CHANGELOG.md](CHANGELOG.md) lists the changes with each
-breaking one marked. The
-[foundation checkpoint](docs/plans/2026-09-26-project-foundation.md) is history: it records
-how projects, materials and Live Review were qualified, and the changelog describes 0.9.0.
-Dated plans and reviews under `docs/` record how earlier work was qualified; the September
-dogfooding and stabilisation work is in [dogfooding](docs/dogfooding.md),
-[headless harnesses](docs/harnesses.md) and Phase 17 of the
-[engine ledger](docs/plans/repo-size-and-r2-migration-2026-09-10.md).
-
-The large scene optimization pass landed in **#131 (`b1ac6ee`)**: Farm hero/shadow
-work, Golden Gate bridge/shadow/reflection work and Foundry planting/MSAA work,
-with count baselines and parity evidence. It did not rewrite saved assets. Optimized
-Farm is live; local Golden Gate `g9-code5` and Foundry `ff3-review2-code5` deliveries
-are now prepared, while production retains its previous seals. #132
-(`ac8456c`) then fixed ORM UV/transform handling. The remaining work below does
-not restart that completed pass.
-
-## Next
-
-The owner's current direction is to stabilize existing core, assets and scenes,
-review behavior, implementation and cycle intent adversarially, and qualify exact
-local artifacts with explicit issues and evidence limits. Troy brainstorming is not
-an adopted specification or task; its construction awaits direct Matt discussion,
-and its pack stays off the public site. No release actions are authorized. The
-[alignment plan](docs/plans/2026-10-02-core-scenes-site-alignment.md) is the current
-work queue and acceptance contract; the [current backlog](docs/backlog.md) tracks
-individual dispositions. The September Farm pilot and October landing plans retain
-their historical scope.
-
-The 3 October local candidate implements ORM isolation, rigid full optimization,
-rebuild versioning, bounded inspection/export diagnostics, compatible dependencies
-and the adopted scene/control repairs. These planned 0.11 changes remain local and
-unreleased. The fresh current-source corpus passes **224/224** at identity
-`58b1b178…`: 86 examples, 113 scene GLBs and 25 saved Troy regression inputs. Every
-per-input result and optimized hash matches the earlier run; 20 compressed terrain
-inputs remain explicitly decoded derivatives. Separate installed-package checks
-qualify the frozen runtime bundles. Root coverage passes 3,187 tests; the complete
-scene-script gate records 108 tests / 615 assertions, and the motion3 portable gate
-passes 713 tests with 18 historical skips. Typecheck and lint pass.
-
-The bounded [stabilization adversarial review](docs/reviews/2026-10-03-stabilization-adversarial-review.md)
-found no new concrete implementation blocker in its reviewed surfaces. Farm pose,
-blend, warming, orbit/pinch and shared first-key ownership repairs pass their focused
-and compiled both-backend checks. Golden Gate water continuity and startup contact
-pipeline preparation have bounded browser/native evidence. These do not establish
-every clipping pose, physical-device behavior, owner appearance or performance.
-Earlier image/count matrices retain their tested build scopes: Farm 160 direct,
-32 paired and 32 settled-count cases; Foundry 120 default images and 56 exact static
-count rows; Golden Gate 78 depth-reference cases per backend. Four shared Farm pilot
-differences and Golden Gate's 17 WebGPU / 11 WebGL2 original strict flags remain.
-
-The code5 local site **`26ecd2b2…`** now qualifies the changed public deliveries:
-564 tests with two historical skips, all 70 Astro files without diagnostics, static
-checks, six NVIDIA control scenarios / 34 checks and nine lifecycle/viewer/failure
-areas pass. Fresh accessibility passes 396 widths and 44 keyboard tours with zero
-issues. Served checks cover 116 downloads, 520 linked files, 89 pages and three
-checksum endpoints. Revision reconciliation covers 292 records / 882 checks with
-zero mismatches; saved model/data/source revisions remain unchanged. The actual
-public Farm diagnostic also passes orbit, pinch and first keyboard movement after
-touch on NVIDIA WebGPU and WebGL2; its earlier warning-classification failure stays
-recorded. Current fresh/managed skills qualify KIT-6 guidance. The final skill-only
-site **`978664da…`** passes fresh static, delivery and HTTP checks. Its strict
-continuity proof establishes 2,721 unchanged product files, including all 200 HTML
-and 243 scene files, against `26ecd2b2…`; earlier browser measurements retain their
-original artifact and scope. Current skill archives match the maintained guidance.
-After documentation freezes, the final offline package byte audit is separately
-recorded at `tmp/alignment-20261003/package-stabilization-final/receipt.json`; the
-qualification ledger records its result. No result is claimed in advance.
-
-Matt selected **focused checks, expanding only if a failure needs investigation**.
-All **18 intended focused runs / 20.5 measured minutes** are collection-valid.
-Raw D41 is **17 pass / 1 fail**, retaining candidate Golden Gate balanced flyover's
-50.1 ms interval at 56.9993 s. The Farm minimal 135-second pair passes D41 with
-original/candidate maxima of 25.0/33.4 ms. Strict post-run quiet checks fail for
-both blocks: CPU maxima of 16.473% and 16.159%, sampled 18.125 s and 6.326 s after
-their respective blocks. The overall focused audit therefore remains **FAIL**.
-One additional unchanged flyover A/B pair completes two measured minutes with
-50.0/16.7 ms maxima, both D41 and all quiet checks passing, with complete cleanup.
-Its earlier 12.694% CPU preflight failure stopped before browser launch and is preserved.
-The candidate's later 50.1 ms miss did not recur; no actionable cause or new runtime
-repair is established. The passing repeat does not erase the original failures.
-All intended focused measurement work is complete; no further benchmark expansion
-is presently justified. The prepared 18-cell / 108-run campaign
-is preserved, deferred and unrun; no all-tier, repeated-run or desktop-speedup claim follows.
-The prior UI3 balanced block stays 7/18 cells and 42/108 runs, with 37 D41 passes and
-five failures; it does not qualify the changed candidate. Golden Gate's separate
-58.3 ms drive event remains distinct from the identified flyover pipeline cause.
-Farm's four allocation observations are complete: neither current observation has
-a natural interval above 40 ms, compared with original intervals of 50.0 and 41.7 ms.
-Current late major GC still takes about 36 ms; source allocation/retainer ownership
-remains unresolved, so no speculative repair or hitch-fixed claim is justified.
-Functional implementation is complete within its tested scope. Strict desktop
-performance acceptance remains open; physical-device and owner appearance review
-are deferred, with owner farmer/water review after the development goals. No desktop
-margin, tablet promotion, full-goal completion, owner approval or deployment is claimed.
-
-1. **Core correctness and upgrades.** Implementation, fresh final-source corpus and
- core/package gates pass. Retain named/LOD, draw/byte and saved-revision contracts
- through final package binding and release review.
-2. **Focused scene follow-ups and asset checks.** Motion3 and code5 functional
- qualification and bounded Farm allocation diagnosis are complete within the recorded
- scope. Intended focused measurements are complete; retain the failed strict
- performance audit and unresolved cause without speculative repairs or further
- benchmark expansion. Do not promote incomplete acceptance evidence.
-3. **Site delivery.** Code5 seals and the exact local launch artifact are qualified.
- Skill-only rebuild and product continuity pass. Record the separate final package
- byte audit after documentation freezes, then review the exact clean artifact before
- any authorized Cloudflare release.
- GitHub validation does not deploy it.
-4. **Troy remains deferred.** Discuss it directly with Matt after the stabilization
- report. Earlier ideas and the [handoff](packs/troy/HANDOFF.md) are reference only,
- not adopted construction/refinement tasks. Keep the pack off the public site.
-5. **Later product milestones.** Ten-pack production, broader importer qualification and
- the v1 public package launch remain separate milestones.
-
-Projects remain optional for every other kind of standalone authoring.
-
-## Disposition of earlier author requests
-
-The former "not adopted in 0.9" list was stale. Current implementations and
-[CHANGELOG.md](CHANGELOG.md) establish these dispositions:
+Updated 5 October 2026, against main `7ed880f` and the published website.
+[CHANGELOG.md](CHANGELOG.md) defines engine release behavior.
+The [backlog](docs/backlog.md) separates completed work, known limits and future decisions.
+Dated plans and reviews retain their original candidate scopes.
+
+## Delivered
+
+Kiln 0.10.0 remains the current engine tag. Standalone authoring, optional projects,
+materials, Live Review, calibrated review lighting, bounded tool results and the
+current CLI/MCP contracts are implemented. Subsequent core fixes and compatible
+dependency upgrades are on main; the planned 0.11 package release is separate.
+
+The completed scene optimization pass landed in #131, ORM UV/transform fixes in
+#132 and the core/stabilization/site follow-ups in #133. Their bounded qualification
+and remaining performance limits are recorded in the
+[3 October release disposition](docs/reviews/2026-10-03-release-disposition.md).
+These are completed changes, not an instruction to restart that pass.
+
+Troy's hub source and in-progress work have been migrated to this computer. The
+maintained, portable scene source is [scenes/packages/troy](scenes/packages/troy/README.md).
+Development continues locally; the hub is used for isolated performance testing.
+The coast, terrain/ocean, fleet transitions, crowds and archers, playable
+Achilles/Hector, combat, blocking, stamina, bots, simple hands/grips and consistent
+scene controls are delivered. The old private playfield-review prototype is retired.
+
+The public site now includes Troy's scene, pack and gallery, together with Farm,
+Vehicles, Bridge and Foundry packs. All five packs have source in `packs/`; all
+four runnable scenes have source in `scenes/packages/`. Vehicles supplies shared
+assets rather than an invented fifth scene. Runtime and Editable downloads are
+available across the packs. Large generated payloads are delivered through
+Cloudflare R2; hydration is optional, so cloning source does not fetch them.
+The [delivery/source plan](docs/plans/2026-10-05-asset-delivery-and-source-organization.md)
+records the adopted organization.
+
+Current site pins at this checkpoint are Farm `r36-local-review`, Bridge
+`g9-code6`, Foundry `ff3-review2-code5` and Troy `troy-20261005-07`.
+The public text encoding cleanup, Troy homepage pictures and revised Foundry
+campus preview are deployed. The current data manifests govern subsequent pins;
+this dated list is not a substitute for inspecting them.
+
+## Next engine and product milestones
+
+1. Release the main-only breaking optimizer changes as the planned 0.11 package,
+ with explicit saved-source replay/migration guidance and installed-package gates.
+ Website deployment did not publish a new engine tag or npm package.
+2. Qualify broader destination/importer support when adopted, especially LOD,
+ visibility and optional compressed runtime derivatives. Current scene delivery
+ does not imply universal Meshopt, Draco or KTX2 loader support.
+3. Develop further packs and the public v1 package as separate milestones.
+ Ten-pack production is not unfinished work in the current Troy release.
+
+The full optimizer now preserves semantic nodes, hierarchy, placements, animated
+targets and LOD relationships; ordinary named meshes can become empty semantic
+nodes. Old unversioned full-mode source must replay with its original engine or
+produce an explicit new revision under the new contract. Standalone saved asset
+revisions are not silently replaced by scene optimization derivatives.
+
+## Qualification limits
+
+The quiet hub target remains 60 FPS at 1440x900, balanced, including combat and
+fleet transitions. Retained Troy timings are scoped to their exact candidates;
+release 07 preserves release 06 model/bank payloads. Frame tails remain, so neither
+mean FPS nor deployment establishes a locked-60 guarantee. Reduced Samsung-tablet
+measurements exist; physical-device and all-tier acceptance remain separate.
+
+The earlier Farm/Bridge/Foundry strict timing failures and unresolved Farm late-GC
+allocation ownership remain in the
+[release disposition](docs/reviews/2026-10-03-release-disposition.md).
+A later passing repeat does not erase those failures. No additional broad
+benchmark campaign or speculative patch follows without a reproducible problem.
+
+Full-army casualties/projectile damage, boarding, wall climbing, an RTS campaign
+and optional living-horse running contacts are outside the delivered Troy endpoint.
+They are potential new scope, not prerequisites for the published scene.
+
+## Authoring request dispositions
| Request | Current state |
| --- | --- |
-| Per-part transforms | Available through requested `listParts` placement output |
-| `userData` and visibility | Bounded JSON extras and `KHR_node_visibility` shipped |
-| Easing and index policy | Cubic animation samplers and indexed/as-built policies shipped; as-built requires optimization off |
-| Sweep self-intersection | Bounded local/adjacent-ring analysis shipped; general distant self-intersections remain outside its guarantee |
-| Hide-node captures | Available in `kiln.capture.v2` |
-| Intentionally open parts | `markOpenShell` shipped; it does not suppress measured overlap failures |
-| LOD authoring and review | `defineLod`, `MSFT_lod`, LOD0 default views and per-level inspection shipped; destination support must still be verified |
-| MCP render output path | Still unadopted; CLI views/capture write files and MCP returns media/receipts |
-
-## Deferred
-
-| Item | State |
-|---|---|
-| `kiln_present` puts up to 16 MiB of base64 in `_meta` | Deferred by decision. The shape is settled and the MCP Apps spec confirms it: a UI iframe may call `resources/read`, so the widget can fetch what the manifest names. The over-limit behaviour is a graceful refusal, not a failure |
-| SEP-2640 / `skill://` resources | Deferred by decision; see ledger 7.12 and 14.1 |
-
-## Dependencies
-
-| Family | State |
-|---|---|
-| Three.js/types and webgpu | Local candidate uses 0.186.1 / 0.186.0 / 0.6.2 with synchronized locks and identities. Native GPU/material/display, fresh 58b1 source corpus, motion3 both-backend controls and code5 site gates pass within their scopes. Intended focused timing is collected: 18 valid runs, raw 17 D41 pass / one fail and two failed post-run quiet gates; a separate unchanged flyover pair passes but does not erase failures. Strict desktop acceptance remains open; the 108-run preparation stays deferred/unrun and owner/device/release boundaries remain |
-| glTF Transform, MCP, material/runtime and site tooling | Compatible updates are installed locally and pass the full core and package gates. Exact versions and evidence are in the alignment plan; production has not changed |
-| Maintainer Node/Bun/npm | Local pins and active hub tools are 22.23.3 / 1.4.2 / 12.2.0; the former Node/npm prefix is retained |
-| TypeScript | Root/scenes use 7.0.2; site and the scene compiler-API alias retain 6.0.3 intentionally |
-| AI SDK/provider/OpenRouter/OpenAI majors | Held by Strands compatibility: installed candidate Strands 1.19.0 requires provider 3 and OpenAI 6. Its Anthropic peer also excludes the newest Anthropic family |
-
-Versions were checked on 2 October and the compatible updates were installed and
-qualified on 3 October 2026. Keep exact release inputs; unresolved scene evidence
-and incompatible major families are not silently accepted by the core test results.
+| Per-part transforms | Requested `listParts` placement output |
+| JSON extras and visibility | Bounded `userData` and `KHR_node_visibility` |
+| Animation easing and index policy | Cubic samplers and indexed/as-built policies; as-built requires optimization off |
+| Sweep self-intersection | Bounded local/adjacent-ring analysis; no general distant-intersection guarantee |
+| Hide-node capture | `kiln.capture.v2` |
+| Intentional open shells | `markOpenShell`; measured overlap failures still apply |
+| LOD authoring/review | `defineLod`, `MSFT_lod`, LOD0 default views and per-level inspection; verify destination support |
+| MCP render file destination | Unadopted; CLI file output and MCP media/receipts remain supported |
+
+Projects remain optional for standalone authoring.
+
+## Deferred decisions and dependencies
+
+| Item | Disposition |
+| --- | --- |
+| Large `kiln_present` metadata | Existing bounded behavior retained; resource-fetch redesign deferred |
+| SEP-2640 / `skill://` resources | Deferred by decision; engine ledger sections 7.12 and 14.1 |
+| Three.js/types/WebGPU | Exact compatible pins 0.186.1 / 0.186.0 / 0.6.2 on main |
+| Maintainer Node/Bun/npm | 22.23.3 / 1.4.2 / 12.2.0; consumer Node support is a separate contract |
+| TypeScript | Root/scenes 7.0.2; site and compiler-API alias 6.0.3 intentionally |
+| AI SDK/provider/OpenRouter/OpenAI majors | Held by Strands peer compatibility; compatible updates are implemented |
+
+Use the current locks, `toolchain.json` and changelog when preparing another release.
+Historical Troy stage notes and the retired review workflow are summarized in the
+[reconciliation record](docs/reviews/2026-10-05-status-reconciliation.md).
diff --git a/docs/backlog.md b/docs/backlog.md
index 03cd722d..7edeb5d6 100644
--- a/docs/backlog.md
+++ b/docs/backlog.md
@@ -1,262 +1,109 @@
# Kiln current backlog
-**Release direction, 3 October:** Matt confirmed the farmer hand fixes and could not
-reproduce the water jump locally. He authorized final performance review, commits,
-main integration and Cloudflare deployment. The focused performance investigation is
-settled for release with its original failures and limits preserved. See the
-[release disposition](reviews/2026-10-03-release-disposition.md). Bridge stills are refreshed; physical-device
-promotion, Troy construction and npm publication remain deferred.
-
-Updated 3 October 2026. This is the maintained queue for the
-[alignment plan](plans/2026-10-02-core-scenes-site-alignment.md). It separates work
-still needed from historical requests already implemented and suggestions awaiting a
-decision. Code and CHANGELOG.md establish shipped behavior; dated reports retain
-their original evidence and limits. No queued item below is an acceptance claim.
-
-Matt's current direction is stabilization of existing core/scenes/assets, with full
-adversarial review of behavior, implementation and cycle intent, bounded known issues,
-and reproducible final-artifact evidence. Report failures, skips, unverified areas,
-risks and blockers explicitly. Troy brainstorming is not a specification, backlog or
-task; construction/refinement awaits direct Matt discussion, and the Troy pack stays
-off the public site. Commit, main integration and site deployment are now authorized; npm publication is not.
-
-## Completed optimization baseline
-
-PR #131 (`b1ac6ee`) landed the large runtime optimization pass in all three scenes.
-Farm's hero/shadow work, Golden Gate's bridge/shadow/reflection work and Foundry's
-planting/MSAA work are implemented. Foundry's 160/160 recorded parity checks passed;
-no Foundry regression caused by that pass has been established. PR #132 (`ac8456c`)
-then fixed engine ORM UV/transform handling, without changing scenes.
-
-Asset GLBs were not rewritten. Farm's optimized runtime is live; production Golden
-Gate g9 and Foundry ff3-review2 still serve their previous sealed runtimes. Local
-`g9-code5` and `ff3-review2-code5` delivery pins are now active in the candidate;
-their publication and remaining qualification are delivery work, not another
-optimization implementation pass.
-
-## Local candidate progress, 3 October
-
-Work is on `codex/core-scenes-alignment`, based on `ac8456c`; a fresh fetch still
-shows zero commits ahead/behind `origin/main`. Changes below are local and
-uncommitted. They have not changed the production website or saved asset revisions.
-The code5 site (`26ecd2b2…`) now has fresh changed-surface qualification. The prior
-code3 site (`a4d1e2d0…`) and UI3 timing retain their original scopes. The bounded
-[adversarial review](reviews/2026-10-03-stabilization-adversarial-review.md) records
-implementation findings and remaining limits. Final skill-only artifact `978664da…`
-passes fresh static/delivery/HTTP checks and strict continuity: all 2,721 product
-files, 200 HTML files and 243 scene files are exact to `26ecd2b2…`; current guidance
-archives are separately verified. After documentation freezes, the final offline
-package byte audit is separately recorded at
-`tmp/alignment-20261003/package-stabilization-final/receipt.json`; the qualification
-ledger records its result. This prose does not pre-claim the audit outcome.
-Matt selected focused timing checks, expanding only for a failure requiring
-investigation; owner farmer/water review follows the development goals.
-All 18 intended focused runs / 20.5 measured minutes are collection-valid, with raw
-D41 17 pass / one fail and two failed post-run quiet gates. A separate unchanged
-two-minute flyover pair passes D41, before/internal/after quiet checks and cleanup;
-it does not erase the original failures. Intended measurements are complete and
-functional implementation is complete within tested scope, but strict desktop
-performance acceptance remains open. No further benchmark expansion is presently justified.
-
-| Items | Implemented locally and checked | Still required |
-| --- | --- | --- |
-| CORE-1 | Shared ORM images isolated per packing pair; serialized, order and other-slot tests; independent adversarial review | Full root/coverage and installed-package gates pass; awaiting release review |
-| CORE-2/3/4 | Rigid merging wired into full; LOD, named pivots, placements and shared-scene boundaries; byte budget, 16-bit splits, precision and failure cleanup; `rigid-v1` rebuild policy | Full root/coverage and installed-package gates pass; release review remains |
-| CORE-5 | Bounded per-anchor draw diagnostics and sampled geometry observations in inspection/export; compact render unchanged | Fresh/managed Codex and Claude setup/upgrade qualification passes; paid dogfood remains separate |
-| ASSET-1 | Consolidated 292 revision/resource rows across all existing families; all 113 scene GLB consumers and 109 auxiliary inputs represented; final code5 reconciliation repeats 882 checks with zero mismatches | 56 Foundry original source/export records are not linked by selected models-only inputs; fresh appearance, functional pivot/clip/scale and owner acceptance remain separate. Bridge previews explicitly apply to the full tier |
-| ASSET-2 | Fresh 224/224 current-source corpus at 58b1 passes: 86 examples, 113 scene GLBs and 25 saved Troy regressions; every result and optimized hash matches the prior run | Structural/deterministic/Khronos scope only; 20 compressed terrain inputs remain decoded derivatives. GPU/material/importer/owner acceptance stays separate |
-| UPGRADE-1 | Compatible geometry, image, MCP/provider and tooling updates; Vite 8.3.2 synchronized; Node 22.23.3/npm 12.2.0 pins and isolated hub tools aligned | Full root/coverage and installed-package gates pass; renderer scene parity remains UPGRADE-2 |
-| UPGRADE-2 | Three 0.186.1/types 0.186.0/WebGPU 0.6.2 coherent; renderer 78 tests, NVIDIA smoke/conformance, core/package and fresh 58b1 corpus pass. Prior bounded scene matrices, motion3 both-backend controls and code5 public/site qualification retain exact scopes; all intended focused measurements are collected | Strict desktop performance acceptance stays open: original 18 runs retain one D41 miss and two failed post-run quiet gates despite a passing flyover repeat. Prior UI3/shared pilot failures, missing latest runtime owner receipt and GG strict flags remain |
-| TOOL-1/2/3, KIT-1 | Exact packs restored and hash-verified; portable browser resolution/profile cleanup; script tests in normal gate; exact-input gate 22 pass with zero skips; WebP serving and actual browser lifecycle checks pass | New release qualification and authorization remain separate |
-| KIT-3, FARM-1 | Morph-weight shadow observation, tested explicit live/invalidation contract, economy trailer exemption; direct images pass 160/160, all six settled X-02 fields match 32/32, and same-phase paired audit finds no introduced failure in 32/32 | Four absolute pilot-image failures are shared with the original; performance and owner acceptance remain separate |
-| KIT-2 | Shared panel bounds/scroll/stacking and nested touch handling; 72/72 DOM contracts. Code5 passes all six NVIDIA public-control scenarios / 34 checks, nine lifecycle/viewer/failure areas and fresh accessibility; skill-only final site has exact product continuity | Physical-device and owner acceptance stay separate |
-| KIT-5 | All 18 intended focused runs / 20.5 measured minutes collected validly; raw 17 D41 pass / one fail. Farm minimal pair passes at 25.0/33.4 ms maxima. Extra unchanged flyover pair passes D41 at 50.0/16.7 ms, all quiet gates and cleanup | Overall focused audit remains FAIL: original candidate flyover 50.1 ms and post-run CPU maxima 16.473% / 16.159% remain. Repeat preflight CPU 12.694% failed before browser and is preserved. No further benchmark expansion presently justified; full 108-run preparation remains deferred/unrun |
-| FARM-2 | Busy-port handling verified; original Farm's 42 files verified; direct 160/160 and settled X-02 32/32 pass; paired 32/32 has no introduced failures; existing r34 B-07 pooling/optimization/ordered collider counts pass for both builds, with 41 cached-shadow textures | Four pilot-image failures are shared, not absolute parity. B-07 retains its r34 visible-source provenance note; no new collider soup/motion oracle. D-64/D-68 adopts the comparison baseline but no hash-bound latest scene-runtime owner acceptance receipt is established, so releaseQualified remains false |
-| FARM-3 | Four valid allocation observations complete; current late GC remains about 36 ms with stable endpoint counts/joints. Separate focused 135-second minimal hero-view pair has original/candidate maxima 25.0/33.4 ms, both raw D41 pass | Post-run CPU maximum 16.159%, sampled 6.326 s after the block, fails strict quiet qualification. Source allocation/retainer ownership stays unresolved; no speculative repair, hitch-fixed, leak-freedom or performance-acceptance claim |
-| GG-1 | Opt-in depth and sky/reflection/bias adaptations; NVIDIA prototype-reference comparison passes 78/78 on each backend; final public/startup/missing-clip probes pass; local `g9-code5` delivery prepared | Authentic original→final retains 17 WebGPU and 11 WebGL2 strict flags; owner review and publication remain separate |
-| GG-2 | Dynamic reflection exclusions and persistent hidden fog banks; 17 focused tests; both-backend 78-view comparisons retain camera/bridge counts and pass against the reconstructed depth reference; local `g9-code5` delivery prepared | Moving-workload timing remains GG-4/PERF-1; owner review and publication remain separate |
-| GG-3 | Verified per-world layout and deterministic 12,706-byte startup projection; code5 public runtime is 1,708,928 raw / 512,680 gzip, within fixed ceilings; both-backend initialization checks pass | Quiet timing remains PERF-1; no tablet tier promotion |
-| GG-4 | Contact-shadow prewarming has focused tests and first-use native proof. Original focused candidate flyover misses D41 at 50.1 ms near 57 s; the nearby 41.3 ms render-call wall interval is not measured CPU. One unchanged extra A/B pair completes at 50.0/16.7 ms maxima with D41, quiet and cleanup passing | The candidate miss did not recur; no actionable cause or new runtime repair is justified. The repeat does not erase the first miss or failed quiet gate. Broader preparation stays deferred; owner review follows development goals |
-| FF-1 | Prior comparison passes 120 default image cases and 56 static X-02 rows. Code5 keeps asset bytes/acceptance; motion3 first-key/follow and code5 controls/lifecycle pass both backends. Selected focused runs have raw D41 passes | Strict desktop performance acceptance remains open because the shared block fails post-run quiet qualification; owner acceptance and historical interior-pose limits stay separate |
-| FARM-4 | Per-render empty-hand correction, borrowed mirrored arm, 150 ms blending and registered warming preserve saved GLBs. Motion3 both-backend entry/start-stop/reentry and code5 public controls pass; bounded adversarial review finds no new concrete blocker | Owner confirmed the hand/animation appearance on the local candidate on 3 October; physical mobile remains a separate scope |
-| GG-5 | Phase/foam-origin fixes and reduced wakes are implemented; bounded shader/render checks pass. Owner could not reproduce the jump in the local candidate on 3 October | Closed for this release on scoped local evidence; reopen with an exact path/device if it recurs. No universal water or physical-device claim |
-| FARM-5 | Desired camera separated from collision projection; motion3 both-backend orbit/pinch and obstacle regressions pass. Code5 public-02 additionally passes real incremental touch orbit/pinch and first W on actual NVIDIA WebGPU/WebGL2 | Physical-device and owner acceptance remain separate. Public-01 behavior passed but its warning-classifier failure stays recorded; public-02 uses the maintained exact classifier |
-| KIT-6 | Pre-#131 joystick cleanup defect fixed with movement ownership; unchanged DOM fixture passes 8/8. Motion3 Farm/Foundry and code5 public Farm first-key/touch checks pass both backends; current fresh/managed guidance copies and final skill-only site continuity pass | The separately recorded final package audit, open strict desktop performance acceptance and deferred physical-device evidence remain separate |
-| SITE-1 | Portable static-only deployment preflight; artifact hashes rechecked before upload; all three upload manifests summarize 486 files with zero problems; actual dry run correctly refuses the dirty checkout | Clean reviewed source/artifact and authorized upload after remaining release work |
-| SITE-2 | Local g9-code5 and ff3-review2-code5 archives reproduce exactly, stage through the normal path and match runtime/page/catalog references; original model/data bytes, saved revisions and checksums are unchanged. Final skill-only artifact continuity passes | Clean authorized R2/Pages delivery/public verification; old objects remain unchanged |
-| SITE-3 | Final skill-only artifact 978664da…: 2,728 files / 199,998,384 bytes; fresh static/delivery/HTTP pass. Strict continuity preserves 2,721 product files, all 200 HTML and 243 scene files from 26ecd2b2…, retaining its 564-test, 70-Astro-file, six-control/34-check, nine-flow, 396-width/44-tour evidence within original scopes. Current skill archives match maintained guidance | Final package audit is separately recorded after documentation freezes; strict desktop performance acceptance stays open and owner/device/release boundaries remain |
-
-The latest portable gate records 713 passes, 18 historical skips and zero failures
-across 123 files; typecheck/lint pass. Core coverage remains 3,187 passes, three
-platform skips, 95.11% functions and 92.32% lines. All six runtime bundles retain
-58b1 identity and installed-package smoke scope. The fresh source corpus at that
-identity passes all 224 inputs with stable before/after source and unchanged assets;
-all per-input outputs equal the prior dd25 run. See
-`tmp/full-optimizer-corpus/current58b1-20261003/qualified.json` and its independent
-`verification.json`. This is source execution, with dist execution covered separately.
-
-Code5 site qualification and final revision/runtime/reference reconciliation are
-complete within the scopes above. Current fresh/managed Codex and Claude copies
-pass after KIT-6 guidance; the final skill-only site rebuild and exact product
-continuity proof pass; the final offline package byte audit is separately recorded
-after these documents freeze. The bounded adversarial
-review reports no new concrete blocker and preserves the authorship/test limits.
-No release or owner approval is implied. The four bounded Farm allocation observations
-are complete, as are the intended focused measurements and bounded flyover repeat.
-The overall focused audit remains failed; strict desktop performance acceptance is
-open, with no further benchmark expansion presently justified. Earlier code3/UI3 receipts and
-the deferred, unrun 108-run preparation stay historical. No all-tier, repeated-run
-or desktop-speedup claim follows from the selected checks.
-
-The prior `f0e49fbc…` site's accessibility receipt freshly checks four widths on the two
-changed documentation routes and reuses 392 widths and all 44 tours from artifact
-`42a41439…` only after exact HTML/dependency identity and unchanged Chrome, axe and
-verifier are proved. That preserved prior receipt combined 60 freshly checked
-routes with 138 byte-identical preliminary routes; its original proof remains
-bound to those checks. That prior receipt and proof are under
-`site/.cache/alignment-release-docs/`; this chain does not qualify activated
-scene/viewer interactions or UI2. Its strict continuation proof rejects reuse for
-UI2 because emitted shared CSS also differs outside the two approved rules; all
-198 routes / 396 widths, 44 tours and 2,138 stops now pass freshly on `a4d1e2d0…`,
-with zero problems and no reused routes. The exact receipt is
-`site/.cache/alignment-site-qualified/accessibility/combined.json`.
-The authenticated Cloudflare list at 10:35 UTC still identifies production
-deployment `33c716c7…` / `b1ac6ee`. A fresh Node-fetch check at 10:50 UTC verifies
-its public build metadata and manifest `5ecbbe76…`, with old `g9` / `ff3-review2`
-pins. The earlier Python-client 403 attempt remains preserved separately.
-No upload or commit has occurred; the successful public recheck is
-`tmp/alignment-20261003/production-recheck-20261003T103132Z/node-public-receipt.json`.
-
-The complete scene-script gate passes **108 tests / 615 assertions / 0 failures**
-across 22 files (`/tmp/kiln-scenes-timing-evidence-scripts-host-final.log`). It covers
-bounded cleanup, measured-interval visibility/focus and actual GPU device-to-canvas
-binding as well as complete readable quiet telemetry. Unknown GPU load or locker
-state cannot qualify as idle. The kit generator reads `scenes/toolchain.json`.
-The quiet-host blocker is resolved without changing thresholds. The balanced
-block completes 7 of 18 cells / 42 of 108 one-minute runs against the exact original
-and UI3 candidate binaries; before, internal and after quiet checks pass. D41 is
-37 pass / 5 fail. All 12 Farm and 18 Foundry runs pass raw D41; the six static
-Foundry fab runs have absent exterior pose fields under the expected schema,
-so these receipts do not claim independently measured interior camera coordinates.
-
-All six Golden Gate flyover runs peak at about 25 s into sampling (about 30 s of
-scene time after warmup), in both builds: A 50.0/66.6/58.4 ms and B 58.2/50.1/49.9 ms.
-Four fail D41; candidate drive pair2 separately fails at 58.3 ms, about 4.408 s
-into sampling. Flyover busy×clock rises +3.45%, +0.05% and +2.91% across the three
-pairs; higher clocks alone do not explain it, but coarse global GPU polling does
-not establish a code cause. No desktop margin or tablet adoption threshold is applied.
-
-Seven economy and four orbit cells were unrun in that historical UI3 campaign.
-The current motion3 campaign prepares all 18 cells / 108 runs and is unstarted;
-the code5 functional/site checks do not qualify its performance. Raw historical
-results remain in
-`scenes/evidence/perf/hub-alignment-2026-10-03/candidate-ui3-balanced/`, with independent
-calculations in `tmp/alignment-20261003/balanced-performance-review/review.json`.
-All five D41 failures and the separate drive event remain recorded. Four bounded
-135-second Farm allocation observations are prepared under
-`scenes/.tmp/alignment-farm/allocation-plan/`, with execution pending. No allocation
-cause, timing pass, deployment or release approval is claimed.
-
-## Work items and required completion evidence
-
-The order is core correctness and compatible upgrades, focused scene follow-ups,
-then qualified site delivery. The rows include regression checks and release chores;
-they are **not a count of unimplemented defects or mandates to rework every asset**.
-The local-progress table above records which implementation and checks already pass. Existing
-accepted evidence remains useful. Recheck the surfaces changed by the next release.
-
-| ID | Work | Required completion evidence |
-| --- | --- | --- |
-| CORE-1 | Fix pre-existing shared-image ORM overwrite | Tests for distinct AO maps, other-slot consumers, order independence and serialized GLB; preserve material/UV/sampler behavior |
-| CORE-2 | Complete deferred engine full-export optimizer | Named pivots, placements, animation targets and all LOD levels survive; caller inventory and public contract settled; separate from landed scene runtime merging |
-| CORE-3 | Correct D2/D3 and bake-frame policy | Actual added bytes per saved draw, conditioned transforms, 65,534-vertex split and end-to-end determinism tests |
-| CORE-4 | Version rebuild behavior and expose diagnostics | Old saved revisions have an explicit replay/migration policy; merge/rejection summary reaches callers |
-| CORE-5 | D23 outputs and maintained skill guidance | Qualified draw estimates and geometry observations; compact output budgets; fresh setup and managed upgrade preserve user customizations |
-| UPGRADE-1 | Compatible dependency and maintainer-tool updates | Peer-compatible exact pins, synchronized locks/toolchain/CI and applicable offline gates |
-| UPGRADE-2 | Renderer-family update | Coordinated Three.js/types/WebGPU dependencies, build identity, GPU smoke and both-backend scene parity |
-| ASSET-1 | Reconcile existing inventories and acceptance | Match final scene consumers, source/material/asset revisions, acceptance states, pack/archive/mirror pins and served download hashes; keep new runtime delivery labels separate from asset revisions. Preserve originals and exclude Troy from the public catalog/artifact |
-| ASSET-2 | Core regression corpus | All 86 examples, exact released scene assets and pinned Troy revisions; engine LOD-aware IO, world-space geometry and semantic assertions |
-| TOOL-1 | Clean-hub exact inputs and browser launch | Public hash-pinned inputs restored; Linux/Windows executable selection; scene builds independent of missing author workspaces |
-| TOOL-2 | Scene test and release coverage | Script tests in normal gate; correct Foundry budget fixture; exact-input gate cannot pass via fixture skips; Golden Gate count baseline |
-| TOOL-3 | Browser process/profile lifecycle | Owned-port handling and cleanup on normal close and launch failure |
-| KIT-1 | WebP static serving (GG-004) | GET, HEAD and range MIME tests with existing access controls preserved |
-| KIT-2 | Consolidate existing HUD/touch workarounds (GG-007, FF-002) | Narrow/short screens, panel scrolling/stacking/focus and nested controls retain behavior; no claim of broken current scene controls |
-| KIT-3 | Shadow motion contract (D26) | Explicit treatment of morphs, shader motion and custom instance attributes; auto detection or documented live-caster behavior tested |
-| KIT-4 | Collision/interaction regression verification | Preserve already-tested doors/joints/proxies when changing optimization; no current broken interaction established |
-| KIT-5 | Shared parity and GPU timing tools (D18, D16) | Thin-line budget and repeat-capture noise; units/provenance for GPU clock and busy-times-clock |
-| FARM-1 | Trailer shadow and baseline work (D8) | Economy trailer exemption with appearance, draw and shadow parity |
-| FARM-2 | Paired pilot audit (D20) | Busy-port fix; new and last-accepted same-phase captures on both backends; texture/count provenance |
-| FARM-3 | Late hero-view hitch (D22) | Desktop GC mechanism reproduced; measure allocations/retainers before a targeted fix and repeated quiet-host validation; retain tractor/joint behavior |
-| GG-1 | Reversed depth (D15) | Opt-in renderer feature, sky/reflection/polygon-offset adjustments, WebGPU/WebGL2 and unsupported fallback evidence |
-| GG-2 | Reflection exclusions (D5/D6) | Traffic, player car and fog banks excluded as adopted; vegetation cards preserved and fog banks stay hidden |
-| GG-3 | Runtime layout delivery (D27) | Verified layout data; missing/corrupt/cancelled-load tests; bundle budget and initialization parity |
-| GG-4 | Shared flyover transition and separate drive hitch (D22) | Startup-context native pipeline proof passes; qualify repeated exact-candidate quiet timing; retain the separate drive event and five historical D41 failures. No tablet tier change |
-| FF-1 | Verify and release the optimized Foundry runtime | Preserve #131's planting work and unchanged interior; reuse existing tests against restored exact inputs; 62 metadata/pin records already agree |
-| FARM-4 | Farmer pose, gesture and transitions | Implemented cadence/relaxed-hand/blend/warming pass focused and both-backend public motion checks; carry them into rebuilt deliveries and affected lifecycle checks. Broader pose/seat and owner review retain their explicit bounds |
-| FARM-5 | Mobile third-person camera | Farm and shared Foundry motion3 checks pass both backends; complete rebuilt delivery checks. Preserve drag distance, pinch in/out and obstruction recovery; physical devices remain a separate boundary |
-| GG-5 | Owner-observed wakes and high-angle water snapping | Reproduce pier wake size/distinctness and water texture/orientation clipping/snapping during high-angle orbit; qualify the targeted visual behavior without wholesale asset rebaking |
-| KIT-6 | Touch-to-keyboard movement handoff | Preserve the newer input when the joystick hides or releases; actual React lifecycle and first-key browser regressions, including cancel/clear ownership |
-| PERF-1 | Focused investigation settled for release | Retain original 50.1 ms miss and post-run quiet failures; targeted unchanged repeat passes. No reproducible actionable regression established, no additional benchmark or speculative patch required. Strict all-run acceptance and 108-minute campaign remain unclaimed |
-| SITE-1 | Portable deploy preflight | Preserve clean-tree/commit/packs/file-limit checks; verify artifact hashes immediately before upload; refusal and dry-run tests |
-| SITE-2 | Prepare affected local scene deliveries | Qualified exact runtimes and producer hashes; synchronized scene-packs, scene-input archives, mirror/upload manifests and applicable catalog scenePack references; unchanged saved asset revisions and old objects |
-| SITE-3 | Final local launch-mode qualification | Full existing assets/scenes, accessibility/touch/keyboard/failure flows and exact served downloads on the final artifact; explicit identity proof for reused evidence; Troy absent. Cloudflare upload/deployment requires a separate authorization |
-| RELEASE-1 | Installed-package and workspace qualification | Full applicable gates, runtime rebuild, package/skills contents and fresh setup/managed-upgrade checks |
-
-CORE-1 can be a small isolated fix. CORE-2 through CORE-4 change full-mode behavior
-and need the planned 0.11 release/migration treatment. A scene runtime optimization
-does not by itself justify replacing standalone asset revisions. Paid harness dogfood
-needs its own concrete authorized run scope; offline qualification stays provider-free.
-
-## Stabilization readiness exit
-
-The final review must connect each adopted change to intended behavior and exact
-source/build/input evidence, reconcile final existing-scene and asset delivery pins,
-and retain a bounded known-issue inventory with impact and disposition. Report local
-failures/skips/unverified areas and external owner/device/release boundaries separately.
-Prior green matrices or site/package checkpoints do not qualify changed bytes without
-an explicit continuity proof. Complete required local gates before reporting readiness;
-no blanket acceptance, release action or Troy task follows from this report.
+Updated 5 October 2026, against main `7ed880f` and the published website.
+This is a work queue and known-limit inventory, not a count of defects.
+The [roadmap](../ROADMAP.md) describes the delivered state and future milestones.
+[Historical reports](reviews/2026-10-05-status-reconciliation.md) keep their original
+test, device and candidate scopes.
+
+## Completed delivery cycle
+
+The hub migration, Troy environment/gameplay/control/asset work, public scene
+integration, pack downloads, portable scene source and source organization are
+complete. Development happens on this computer; the hub is reserved for isolated
+performance testing. The canonical Troy entry is
+[scenes/packages/troy](../scenes/packages/troy/README.md).
+
+All five packs (Farm, Vehicles, Bridge, Foundry and Troy) have checked-in editable
+source; all four runnable scenes have maintained source packages. Runtime and
+Editable downloads are public. Large generated files remain outside the normal
+clone through hash-pinned Cloudflare R2 delivery and explicit hydration.
+Bridge is in the gallery. The unwanted Battle before the gate composition is
+removed from gallery presentation while its saved source/history remains.
+
+Current scene pins are Farm `r36-local-review`, Bridge `g9-code6`, Foundry
+`ff3-review2-code5` and Troy `troy-20261005-07`.
+The horse-ear child revision, loading transport/bundling changes, public-text
+encoding guard, Troy homepage images and revised Foundry campus preview are delivered.
+The obsolete private `/reviews/troy/` prototype has been retired; use the normal
+public scene, pack and gallery flow.
+
+The following original alignment IDs are closed within their recorded scope.
+They remain useful regression contracts when a later change touches that surface.
+
+| IDs | Implemented and delivered scope |
+| --- | --- |
+| CORE-1 | Shared-image ORM isolation, distinct AO consumers and serialized output |
+| CORE-2, CORE-3, CORE-4 | Full-export semantic preservation, draw/byte policy, conditioned transforms, split limits and versioned rebuild diagnostics on main |
+| CORE-5 | Bounded D23 diagnostics and maintained guidance, fresh setup and managed upgrades |
+| UPGRADE-1, UPGRADE-2 | Compatible exact dependencies/toolchain, renderer identity and scoped both-backend checks |
+| ASSET-1, ASSET-2 | Revision/inventory reconciliation, source corpus and hash-pinned public deliveries; Troy is now included |
+| TOOL-1, TOOL-2, TOOL-3 | Portable inputs/browser launch, script gates and owned browser lifecycle |
+| KIT-1, KIT-2, KIT-3, KIT-4, KIT-5, KIT-6 | WebP serving, HUD/touch/input ownership, shadow motion, preserved interaction and parity/timing tools |
+| FARM-1, FARM-2, FARM-4, FARM-5 | Trailer shadows, paired pilot audit, farmer motion and scoped camera/input repairs |
+| GG-1, GG-2, GG-3, GG-4, GG-5 | Depth/reflection/layout delivery, targeted startup pipeline repair and owner-observed water fixes; timing limitations remain below |
+| FF-1 | Optimized Foundry runtime with retained campus/interior behavior |
+| PERF-1 | Intended focused investigation completed; failed strict results preserved, no actionable new regression established |
+| SITE-1, SITE-2, SITE-3 | Portable deploy preflight, sealed deliveries and qualified public flows, including Troy |
+| RELEASE-1 | Scoped installed-package/workspace qualification; a new tagged package release is separate |
+
+PR #131 completed the scene runtime optimization pass. #132 fixed ORM UV/transform
+handling; CORE-1's shared-image isolation and the remaining core changes landed
+through #133. The v0.10.0 tag predates those later changes. Main/site delivery does
+not publish an npm release or retroactively qualify new package bytes.
+
+## Remaining package and product work
+
+| Item | Completion boundary |
+| --- | --- |
+| Planned 0.11 engine package | Tag/package release with breaking full-mode contract, original-engine replay versus explicit child-revision migration, exact runtime/skills bytes and installed-package gates |
+| Public v1 and further packs | Separate product milestones; ten-pack production is not a Troy release chore |
+| Broader importer qualification | Actual destination tests for LOD, visibility and supported export profiles |
+| Optional compressed runtime derivatives | Explicit loader/profile adoption and measured size/startup/render tradeoffs; current shared runtime loading is not universal Meshopt/Draco/KTX2 support |
+| New derivative guidance | Demonstrated authoring/workflow evidence beyond Troy before promoting a scene-specific lesson to shared guidance |
-## Implemented or superseded requests
+Full optimization preserves semantic nodes while potentially emptying ordinary
+named mesh geometry. Off/palette modes retain their documented geometry contracts.
+Old unversioned full-mode source must use the original engine or an explicit new
+revision; never silently rewrite saved revisions.
-- PR #132 shipped UV/transform guards and other recorded fixes. It did **not** fix
- CORE-1's shared-image overwrite.
-- PR #131 landed the completed scene runtime optimization work. It did not
- implement every adopted follow-up or reseal all production runtimes.
-- The maintained compose-scene skill includes #131's optimization lessons; the
- published website archive matched main at the 3 October audit. Four narrow
- reference clarifications are now local, pending release. The v0.10.0 tag predates
- #131, and Troy's current workspace has no optional compose skill installed.
-- FARM-001 through FARM-007, GG-001/FF-001 and GG-009/GG-010 are recorded as done.
-- FF-003's callback-cancellation identity handling is present in the shared kit and
- tests; it is no longer an open implementation request. Full failure-flow acceptance
- remains part of each new scene release.
-- Requested placement output, extras/visibility, cubic easing, index policy,
- hide-node captures, intentional open shells and LOD support shipped. Sweep analysis
- has bounded local coverage, not a general distant-intersection guarantee.
-- Foundry's earlier deferred-startup repair and Golden Gate's earlier rejected bridge
- variant are superseded by their accepted revisions; do not reopen them from old logs.
+## Known limits and evidence-triggered work
-## Held, optional or awaiting evidence
+| Item | Current disposition |
+| --- | --- |
+| FARM-3 late hero-view hitch | Late major GC and allocator/retainer ownership remain unresolved. Diagnose a reproduced event before proposing a repair; no hitch-fixed claim |
+| Farm/Bridge/Foundry strict timing | Retain original failed D41 intervals and post-run quiet checks. Passing repeats do not erase failures. The prepared 108-run campaign remains deferred and unrun |
+| Troy frame tails | Quiet-hub candidate means are scoped evidence, not locked 60 FPS. Target stays 1440x900 balanced, including combat/fleet transitions |
+| Device/tier acceptance | Samsung reduced-tier measurements exist; they do not establish tablet 60 FPS or universal mobile qualification |
+| Xclipse classification (GG-008) | Requires actual device strings and measured behavior |
+| Sweep analysis | Bounded local/adjacent coverage; distant self-intersections are outside the guarantee |
+| Foundry S3 cross-pass and roof/seam design | Pre-existing asset/design questions; retain campus structures pending a specific design decision |
+| Full-army damage, boarding, wall climbing, RTS and living-horse contacts | Outside the accepted Troy endpoint; require new scope |
+
+The [3 October release disposition](reviews/2026-10-03-release-disposition.md)
+retains exact performance failures and source-scoped conclusions. The dated Troy
+reports likewise describe their tested stages, not the latest runtime pin.
+No new benchmark campaign, paid provider run or wholesale rebake is required
+solely because a historical report used pending-release language.
+
+## Optional or deferred decisions
| Item | Disposition |
| --- | --- |
-| Troy scene ideas | Brainstorming only, not an adopted specification, backlog or task. Construction is deferred for direct discussion with Matt; foundation exit does not start it automatically |
-| Troy pack/refinement/publication | Keep off the public site. No refinement or publication task is adopted; existing revisions may remain local regression inputs without changing this boundary |
-| AI SDK/provider/OpenRouter/OpenAI major upgrades | Blocked by current Strands peer compatibility; compatible-family updates are locally qualified |
-| Site TypeScript 7 | Keep TypeScript 6 while Astro's checker/compiler-API consumers require it |
-| Xclipse tier classification (GG-008) | Needs actual device strings and performance evidence |
-| FARM-008 instancing/warm-up advice | Measure applicability per scene; not a proven global defect or automatic rewrite |
-| Shared meshopt loader (GG-002) | Optional consolidation; Golden Gate has a working local path |
-| Terrain derivative encoder (GG-005) | Regeneration tooling is distinct from the usable sealed terrain; decide when new tiles are needed |
-| Static scene export (GG-006) | New delivery contract requiring a separate decision |
-| Foundry S3 cross-pass and roof/seam design | Pre-existing asset/design questions, not #131 regressions; D-75 retains campus structures. Change only after an explicit design disposition |
-| Next vegetation cycle (D33) | Adopted later performance exploration: per-tree culling before LOD, with visual/GPU-work gates. Keep separate from shipping the completed pass |
-| R8 shadow targets, reflection cadence, output-pass depth and extra merge layouts | Optional measured follow-ups; not silently prerequisites |
-| MCP render file destination | Unadopted; existing CLI file output and MCP media/receipts remain supported |
-| SEP-2640 resources and large present metadata | Retain the roadmap's recorded defer decisions |
-| Ten-pack production, broader importers, public v1 package | Later milestones with their own acceptance |
-
-Update this ledger when evidence or an owner decision changes a disposition. Link the
-test/report/revision that closes an item. Historical reports should gain a current
-pointer where needed, not be rewritten to claim new acceptance.
+| AI SDK/provider/OpenRouter/OpenAI majors | Strands peer compatibility holds these upgrades; compatible-family updates are on main |
+| Site TypeScript 7 | Retain TypeScript 6 for Astro checker/compiler-API compatibility |
+| FARM-008 instancing/warmup | Measure applicability per scene; not a proven global defect |
+| Shared Meshopt loader (GG-002) | Optional consolidation of the existing Bridge-specific path |
+| Terrain derivative encoder (GG-005) | Regeneration tooling when new terrain tiles are needed |
+| Static scene export (GG-006) | Separate delivery-contract decision |
+| Vegetation cycle (D33) | Later per-tree culling exploration with visual and measured GPU-work checks |
+| R8 shadows, reflection cadence, output depth, extra merging | Optional measured follow-ups |
+| MCP render file destination | Unadopted; existing CLI output and MCP media/receipts remain valid |
+| SEP-2640 resources and large present metadata | Recorded deferrals; see the roadmap and engine ledger |
+
+## Maintaining this queue
+
+For a changed surface, connect implementation to focused tests, exact source/input
+identity and the affected public or installed-package flow. Preserve failures,
+skips, owner/device limits and dated evidence. Completed migration and publication
+do not reopen from old logs; genuinely new scope needs its own completion boundary.
diff --git a/docs/reviews/2026-10-05-status-reconciliation.md b/docs/reviews/2026-10-05-status-reconciliation.md
new file mode 100644
index 00000000..5cf6acbc
--- /dev/null
+++ b/docs/reviews/2026-10-05-status-reconciliation.md
@@ -0,0 +1,42 @@
+# Status reconciliation, 5 October 2026
+
+This record reconciles the living roadmap/backlog and Troy handoff against main
+`7ed880f`, after the public scene, portable source, dual download profiles,
+startup transport, text repair and Foundry preview releases.
+
+The earlier uncommitted roadmap/backlog contained accumulated Troy stages,
+intermediate population/hand/shield proposals, paused-goal messages and
+pre-publication tasks. Those statements described their checkpoints. They no
+longer define current work. The living documents now close migration, Troy
+production, source organization, download integration and publication while
+retaining genuine package milestones and scoped performance limitations.
+
+The stage-specific records remain intact, including:
+
+- [Weapon contact](2026-10-04-troy-weapon-contact.md).
+- [Fleet, blocking and hands](2026-10-05-troy-fleet-block-hands.md).
+- [City ground and gate](2026-10-05-troy-city-ground-gate.md).
+- [Startup comparison](2026-10-05-troy-startup-phase-comparison.md).
+- [Controls, shields and tablet](2026-10-05-troy-controls-shield-tablet.md).
+- [Pre-publication completion checkpoint](2026-10-05-troy-completion-status.md).
+- [Existing-scene release disposition](2026-10-03-release-disposition.md).
+- [Text and preview release](2026-10-05-site-text-and-preview-release.md).
+
+Original uncommitted document bytes and the retired prototype were hash-verified
+into the local ignored `site/.localdata/docs-reconciliation-20261005/before/`
+snapshot before reconciliation. Existing recovery archives, saved asset revisions
+and ignored audit receipts were left intact.
+
+The five untracked prototype files had no maintained callers: the private stager,
+its tests, review-session module/test and `site/src/pages/reviews/troy.astro`.
+They are retired in favor of the normal public `/scenes/troy/` flow and portable
+source package. The historical delivery plan's ignored review receipts are evidence,
+not a dependency on those files.
+
+The living Troy status/handoff now lead with runtime 07 and portable scene source.
+The dated completion report gains a historical-status pointer; its original
+stage-36/38 measurements and pending-release account remain unchanged.
+
+This reconciliation changes documentation and removes an unused local prototype.
+It does not change scene bytes, promote hardware acceptance, publish an engine
+package, erase failed performance runs or authorize a new benchmark campaign.
diff --git a/docs/reviews/2026-10-05-troy-completion-status.md b/docs/reviews/2026-10-05-troy-completion-status.md
index a017e2a4..85b2f67b 100644
--- a/docs/reviews/2026-10-05-troy-completion-status.md
+++ b/docs/reviews/2026-10-05-troy-completion-status.md
@@ -1,5 +1,10 @@
# Kiln and Troy completion status, 5 October 2026
+Historical pre-publication checkpoint: the stage-36/38 measurements and pending
+release statements below retain their original scope. Troy is now public at
+runtime 07; use the [maintained status](../../packs/troy/STATUS.md) and
+[reconciliation record](2026-10-05-status-reconciliation.md) for current disposition.
+
The hub work and in-progress source are on this computer. The current scene is source stage 38; the public artifact is `troy-20261005-02`. This computer owns development and asset authoring. The hub is reserved for isolated performance testing.
Completed: connected coast/city terrain and ocean, procedural sand/rock detail, fleet arrival/unloading/formations, 48 archers with corrected poses, simple hands/grips, differentiated playable heroes, light/heavy attacks, stamina, blocking and bots. Shield rest is at the side; blocking raises it in front with the forearm across the inner plate. A/D strafe correctly. Left/right mouse attacks supplement J/K and touch. Primary controls remain visible; secondary controls start folded, as in Farm/Fab/Bridge.
diff --git a/packs/troy/HANDOFF.md b/packs/troy/HANDOFF.md
index 051dd4b6..b99c5a98 100644
--- a/packs/troy/HANDOFF.md
+++ b/packs/troy/HANDOFF.md
@@ -1,9 +1,27 @@
# Continue Troy locally
-Scene source: `C:/Users/Mattm/X/kiln-dogfood/v1-readiness-2026-10/review/troy-expansion/scene`. Character authoring workspace: its sibling `troy-character-repair-2026-10-04/ws`; original authoring revisions remain in `troy-refine/ws`. Engine and public site: this repository.
+The maintained scene source is [scenes/packages/troy](../../scenes/packages/troy/README.md)
+in this repository. Follow its clean-checkout installation, `hydrate` and `dev`
+instructions. Edit `web/` locally; the preview serves maintained source first and
+verified GLBs/banks from the ignored runtime-data cache. No private author workspace
+or Kiln server is needed for scene development. The current runtime pin is
+`troy-20261005-07`; a changed pin requires rehydration.
-Start with `review/troy-expansion/LOCAL-START.md` in the private repository. Its platform helper selects the local engine, pinned Node and browser. Author assets in a separate workspace; keep controllers, terrain and optimization derivatives in the scene. Current banks are infantry-v5 and archer-v9, with exact master pins in `scene/web/inputs.json`.
+Asset authoring and rebaking are separate tasks. Use a live Kiln asset workspace
+and the exact saved source/material inputs supplied by the Editable downloads and
+pack resources. The original private `troy-expansion/scene`,
+`troy-character-repair-2026-10-04/ws` and `troy-refine/ws` locations preserve authoring
+history; they are not dependencies for running the portable public scene.
+Do not modify historical bank inputs without rebaking and verifying their hashes.
-Public routes are `/scenes/troy/`, `/packs/troy/` and `/gallery/troy/<slug>/`. The site build stages the exact sealed download from `site/src/data/troy-delivery.json`, without an author's absolute workspace path. Downloads are CC0 assets and MIT browser scene code. Captures are actual scene frames and master GLB renders.
+Public routes are `/scenes/troy/`, `/packs/troy/` and `/gallery/troy/<slug>/`.
+The site stages the sealed artifact in `site/src/data/troy-delivery.json`.
+Assets are CC0 and browser scene code is MIT. Captures are actual scene frames and
+GLB renders. Use the normal scene preview and downloads; the private
+`/reviews/troy/` prototype is retired.
-Read `STATUS.md` and the engine's `docs/reviews/2026-10-05-troy-completion-status.md` for current evidence and limits. Keep the hub quiet before timing an identified candidate. Existing recovery parents, rejected candidates and outputs remain preserved; cleanup belongs to the separate cleanup task.
+[STATUS.md](STATUS.md) and the [current backlog](../../docs/backlog.md) describe the
+delivered scope and known limits. Dated reviews retain their original candidate
+identity. Keep the hub quiet before measuring an exact candidate at 1440x900 balanced.
+This computer owns source and development. Recovery parents, rejected candidates
+and saved revisions remain preserved; broad cleanup is a separate task.
diff --git a/packs/troy/STATUS.md b/packs/troy/STATUS.md
index 698941aa..55a59107 100644
--- a/packs/troy/STATUS.md
+++ b/packs/troy/STATUS.md
@@ -1,13 +1,40 @@
-# Troy status, 5 October 2026
+# Troy status
-The hub work and in-progress source are on this computer. The current scene is source stage 38; the public artifact is `troy-20261005-02`. This computer owns development and asset authoring. The hub is reserved for isolated performance testing.
+Updated 5 October 2026, against main `7ed880f`.
-Completed: connected coast/city terrain and ocean, procedural sand/rock detail, fleet arrival/unloading/formations, 48 archers with corrected poses, simple hands/grips, differentiated playable heroes, light/heavy attacks, stamina, blocking and bots. Shield rest is at the side; blocking raises it in front with the forearm across the inner plate. A/D strafe correctly. Left/right mouse attacks supplement J/K and touch. Primary controls remain visible; secondary controls start folded, as in Farm/Fab/Bridge.
+Troy is published at `/scenes/troy/`, `/packs/troy/` and its normal gallery pages.
+The current runtime pin is `troy-20261005-07`. The maintained source is
+[scenes/packages/troy](../../scenes/packages/troy/README.md), with an explicit
+runtime-data hydration step for a clean checkout. Development is local; the hub
+is reserved for isolated performance testing.
-Stage 38 passes 265 scene tests, 14 actual browser input cases, and three public open/play/Exit cycles on each graphics backend with complete resource release. Twenty-one assets have normal gallery pages and 3D previews, CC0 downloads and actual captured pictures. The public scene and pack follow the other website entries.
+Delivered: connected coast/city terrain and ocean, procedural sand/rock detail,
+fleet arrival/unloading/formations, 720 ordinary actors and two heroes, corrected
+48-archer poses, simple hands/grips, differentiated playable Achilles/Hector,
+light/heavy attacks, stamina, blocking and bots. Shield rest is lateral; blocking
+raises it in front with the forearm across the inner plate. A/D strafe correctly.
+Mouse attacks supplement J/K and touch. Primary controls remain visible and
+secondary controls start folded consistently with the other Kiln scenes.
-The retained stage-36 quiet hub matrix measured 52 windows at 1440x900 balanced. All WebGPU case means met 60 FPS; its slowest mean was 62.26 FPS. WebGL2 later unloading measured 59.46–59.93 FPS. Frame tails remain above 16.7 ms, so this is not a locked-60 claim. GPU timestamps were unavailable. Samsung tablet minimal mode measured about 38–56 FPS. These timings qualify stage 36, not the later input/pose edits in stage 38.
+The public pack has CC0 Runtime and Editable asset downloads, saved source/revision
+resources and normal gallery previews. Browser scene code is MIT. Source packaging,
+startup transport/bundling, the horse-ear child revision, public text cleanup and
+the selected scene pictures have shipped. The private playfield-review prototype
+is superseded and retired.
-Engine typecheck, lint and focused checks pass. Broad local Windows runs retain one asset-import EPERM failure; two site fixtures cannot create file symlinks here. Their assertions remain intact. Exact-commit CI and the clean production build govern release readiness.
+Performance acceptance targets 60 FPS on the quiet hub at 1440x900 balanced,
+including combat and fleet transitions. Exact-candidate checks are retained;
+release 07 corrects text while preserving release 06 model/bank payloads.
+Frame tails remain, so this is not a locked-60 guarantee. Samsung reduced-tier
+measurements do not establish tablet 60 FPS. Earlier stage-36/38 results retain
+their original source scope in the
+[historical completion report](../../docs/reviews/2026-10-05-troy-completion-status.md).
-The endpoint is this environment and two-hero interaction. Full army casualties, projectile damage, boarding, wall climbing and an RTS campaign are outside it. The optional horse is supplied as an asset; its running-contact work remains outside the active scene.
+The delivered endpoint is this environment and two-hero interaction. Full army
+casualties/projectile damage, boarding, wall climbing, an RTS campaign and optional
+living-horse running contacts are outside it. The [current backlog](../../docs/backlog.md)
+records separate engine/package milestones and known limits.
+
+Use the current runtime and delivery manifests for subsequent changes.
+[HANDOFF.md](HANDOFF.md) gives the local continuation path; dated reviews and recovery
+archives preserve history rather than redefining the latest source.