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.
Scenes build report — M1 released; M2 closed as the record (D-20); M3 Farm features delivered; M4 Farm hardening delivered, hub runs pending
Historical report from the original scene workspace. Evidence paths below refer to that workspace's retained local captures, which are not shipped in this repo. Use README.md for current source, build and staging instructions.
Date: 2026-09-30 (M4; earlier sections dated 2026-09-29 keep their dates). Workspace: C:/Users/Mattm/X/kiln-commons/scenes/. Branch remains work/scenes-v1.
M0 passed its continuation gate and M1 is released at KIT-M1: READY 2026-09-29. M2 (Farm parity) is closed as the record under owner direction D-20: parity with the sealed pilot is the baseline, and M2 closes on the evidence gathered, without tightening thresholds or chasing residual differences. M2a (staging, static world) and M2b (optimization stack; 24 of 24 static view/backend pairs) pass. M2c (collision, walking, doors, tractor, follow camera, HUD) passes a 4,291-frame lockstep against the sealed play.mjs and B-08 on both backends. M2d records the play positions, the stream crop, the r34 staging run, the Farm contract suite on both releases and profiles, bundle sizes and the hub run kit. P-16 and P-19 were deferred with reasons at that close.
M3 (round 5, 2026-09-29, TASK-M3.md) delivers the Farm's new features; P-31 to P-35 pass. M3a: mobile play made for touch under D-22 (joystick, drag look, pinch zoom, one context tap target, no button pads), B-09 on emulated touch and with real touch on the owner's Galaxy Tab S9 FE, the site's 390 px HUD overlap fixed. M3b: the Farm tier table and a grass-only live ladder (R5-01), B-10 through an adapter-string hook, B-13, P-16. M3c: public UI, accessibility, pause and reduced motion on the new outputs (B-11, B-12, P-34, P-35). Outputs rebuilt under the label m3; a device kit; tablet runs (tier from the real adapter strings, a governor soak, WebGL2 play positions, frame time per look option); a dev-build look exploration with one contact sheet and the D-18 ACES/Neutral captures, nothing made a default; one bounded X-04 attempt. P-19 stays dropped. The full CPU suite caught one M3 regression (U-18: the D-22 HUD edit had reworded the touch tractor status that SPEC 13.1 keeps verbatim); the pilot text is restored, the outputs were rebuilt, and every counted functional run was repeated on the final build (desktop label m3-final, tablet suffix -final). The hub kit's look-runner plumbing check found a second defect: the bundled static server opened a stray listener on 4400 and kept the runner alive (FARM-005); the guard is fixed with a regression test, and the kit is rebuilt and rechecked. An owner report of WebGL2 hitches on this PC (20:35) is queued as the first M4 item; M3 records observations only, and the frame measurements now keep every interval and count frames over 100 ms. The coordinator's hub run of the r34 kit (17:25 to 19:08 UTC) passes X-01 on both backends and X-04 on WebGL2 and fails X-04 on WebGPU; the tablet at its assigned tier misses the SPEC 20.2 tablet proposal (X-07). Both are owner questions.
M4 (round 6, 2026-09-29 21:43 EDT to 2026-09-30 00:52 EDT, TASK-M4.md with the owner's decisions of 22:05 and 22:10) hardens the Farm; the hub runs are the coordinator's. The owner's WebGL2 hitches on minimal are found and fixed: three 0.186.0 compiled a program for each instanced mesh lazily, when it first came into view (a synchronous link of about 100 ms on WebGL2), and uploaded every per-mesh instance buffer each frame. The Farm now compiles every drawable before ready (the SPEC 6.4 warm pass the owner allowed at 22:05, D-23, bounded to 8 chains at a time), reads instance matrices as instanced attributes and uploads a batch only when it moved (FARM-006), with no look change (the A/B passes). On this PC (indicative, with load samples) WebGL2 minimal went from 6 frames over 100 ms in three orbits to none, and nothing is compiled after ready on either backend. A hitch rule in X-05's shape is proposed for the coordinator (per 60 s run at every tier: 0 frames over 100 ms, 0 over 50 ms, longest at most 50 ms). X-04 is now the time to the first stable frame for both scenes (D-23; SPEC 6.4 and 5.2(3) amended). FARM-004 (the governor evaluates on 60 samples or a spanned window) and FARM-007 (the Tab S9 FE class starts on minimal on WebGPU, owner 22:10, D-25) are done with tests; on the tablet X-07 now passes SPEC 20.2's row at minimal (about 60 FPS, p95 17.2 to 17.5 ms), and forced high steps down and locks within X-10. X-11 and X-12 pass. The outputs are rebuilt under m4 (public chunk 1,581,284 B / 462,421 B gzip, 91.30 % / 91.36 % of the Farm ceiling), every counted functional run passes on them (contract suite on both profiles, B-08, B-09 emulated, B-10 and B-13), the hub kit is rebuilt with X-04's new moment and X-05 as SPEC asks, and the scratch outputs M3 left are deleted.
Owner rules of 2026-09-29 (11:35 and 11:40): timing on this PC is dirty while other agents run here (at the time four Claude sessions, two Kiln render services and Playwright), so a PC timing carries its load sample, is labelled "indicative, PC under load" and never settles a budget, threshold or look; browser checks on this PC run headless, and a headed window is pinned to the primary display and recorded. No timing from this PC is cited in this report. Every parity, contract and B-08 run was headless (the kit's launchChrome default). The only headed runs were the two hub-kit smoke attempts (11:32 and 11:38; the second, a plumbing check, started after the 11:35 rule) and the functional diagnostics .tmp/diag-tractor.mjs and .tmp/diag-tiers.mjs (now headless), with 1920 × 1080 windows at Chrome's default position, which the owner reports open partly on the second display. Their timing values are invalid and were discarded unread; the smoke logs record the load (CPU total 1.1–5.4%; nvidia-smi utilization 20% and 46%, not the Windows 3D-engine counter). Pixel comparisons (B-06) do not depend on load. In M3 and M4, PC timing appears only where the task asks for it (in M3 the look options' frame time and the X-04 attempt; in M4 the hitch counts before and after the fix, the warm-pass concurrency A/B and the X-04 first-stable-frame samples), each labelled indicative with its load sample beside it, and it settles nothing. Frame time on the tablet is the tablet's own and is cited with its device state; the tablet's time to ready is indicative only, with this PC's load sample. The hub's numbers are the coordinator's run.
Known pinned-infrastructure limitation under C-03: React DOM 19.3.0 installs one shared nonpassive document selectionchange listener (react-dom-client.production.js:13824), and its unmount path (:18499) retains it. The cold-mount audit and every qualified teardown identify React’s true _reactListening* marker and leave exactly that one listener, with no other document/window listener or count growth. No supported RootOptions fix exists; no mounting architecture, dependency patch or document-method interception was introduced. Cold proof, exact source locations and rejected alternatives.
Milestones
| Milestone | State | Date |
|---|---|---|
| M0 | Complete continuation gate; B-00e fails with required Z4 fallback | 2026-09-29 |
| M1 | PASS; kit API released | 2026-09-29 |
| M2 | Closed as the record under D-20. M2a PASS (staging, units, 24 static pairs, woodland A/B); M2b PASS (24/24 static pairs with X-02 and B-07 goldens); M2c PASS (oracle lockstep, B-08 24/24 on both backends); M2d recorded (play positions on WebGPU, stream crop, r34 staging run, contract suite r33/r34 on both profiles, bundles, hub run kit). Deferred: P-16 (B-10 is M3b), P-19 (cut line, D-09), B-13 (M3b), X-01/X-04 (hub timing) | 2026-09-29 |
| M3 | PASS: P-31 to P-35. M3a B-09 (emulated 16/16 per backend; tablet with real touch 14/14 per backend, two emulation-only steps skipped); M3b B-10, B-13, P-16, U-15, U-16; M3c B-11, B-12, P-34, P-35 on the m3 outputs; B-08 on m3; device kit and tablet runs; look contact sheet and D-18 A/B captures; X-04 attempt. X-01 passes and X-04 fails on WebGPU on the hub (coordinator's run); X-07 fails the SPEC 20.2 tablet proposal; X-10 passes on the tablet. P-19 stays dropped |
2026-09-29 |
| M4 | Builder's part delivered (round 6): hitch cause found and fixed (warm pass, instanced attributes, FARM-006), hitch rule proposed; X-04 redefined as the first stable frame (D-23) and measured locally; FARM-004 and FARM-007 with tests; on the tablet X-07 passes at minimal and X-10 holds with FARM-004 acting; X-11 and X-12 pass; m4 outputs with the contract suite, B-08, B-09 (emulated), B-10 and B-13 passing; hub kit rebuilt; M3 scratch outputs deleted. Deferred to the coordinator's hub run: X-01 on m4, X-04, X-05, X-06; X-08 and X-09 deferred (no phone); B-14's Farm consumer is the site round's |
2026-09-30 |
The recorded release line is KIT-M1: READY 2026-09-29. Golden Gate consumes the frozen kit API; every kit change since is a compatible addition (FARM-001's optional preparation hooks; cpuRenderTimes() in the test API for the hub kit; in M3 the launchChrome window size (FARM-002), the test-only adapter-string hook (FARM-003), the chase rig's pinch zoom and pinch-drag guard (GG-010) and the static server's entry guard (FARM-005); in M4 the governor's elapsed-window rule (FARM-004), dynamic batches written only when a source moved (FARM-006) and the X-07 classification (FARM-007), each entered in KIT-REQUESTS.md; no exported name or signature changed).
Build, stage, serve and test
The workspace, lockfile, exact pins, kit modules, generated demo pack and public/test/dev outputs are qualified. Run from the workspace root; the wrapper selects the installed Node 22.23.2 and workspace npm 12.0.2 process-locally, with Bun 1.4.2.
& ./scripts/toolchain-run.ps1 run check:toolchain
& ./scripts/toolchain-run.ps1 run check:pins
& ./scripts/toolchain-run.ps1 run typecheck
& ./scripts/toolchain-run.ps1 test packages/farm packages/scene-kit scripts
& ./scripts/toolchain-run.ps1 run demo:build
& ./scripts/toolchain-run.ps1 run check:hygiene
& ./scripts/toolchain-run.ps1 run check:licenses
& ./scripts/toolchain-run.ps1 run check:site-compat
& ./scripts/toolchain-run.ps1 run demo:serve
demo:build stages the five-file, 4,699-byte synthetic pack and builds all three kit demo outputs; demo:serve verifies and serves packages/scene-kit/dist/standalone/ at http://127.0.0.1:4400/ (stop with Ctrl+C). The unit command names this builder's packages so that the second builder's packages/golden-gate tests stay theirs. M1 qualification: Profile 1, Profile 2, gate.
Farm, from the same root (staging reads the sealed archives from the local mirror):
$env:KILN_ASSET_MIRROR = 'C:/Users/Mattm/X/kiln-commons/site-build/mirror'
& ./scripts/toolchain-run.ps1 run stage:farm -- --release r33
& ./scripts/toolchain-run.ps1 run stage:farm -- --release r34 --check-site-manifest C:/Users/Mattm/X/kiln-site-workbench/site/src/data/upload-manifest.json
& ./scripts/toolchain-run.ps1 scripts/build-farm.ts --release r33 --all
& ./scripts/toolchain-run.ps1 scripts/build-farm.ts --release r34 --all --out-root packages/farm/dist/r34
& ./scripts/toolchain-run.ps1 scripts/run-farm-contract.ts --release r34 # add --webgl for forced WebGL2
& ./scripts/toolchain-run.ps1 scripts/capture-farm-parity.ts --label <new> --release r34 --stage m2d --new-root packages/farm/dist/r34/test --views 'house-interior,house-porch,house-window-out,play-house-door'
& ./scripts/toolchain-run.ps1 scripts/run-farm-play.ts --label <new> --backend both
& ./scripts/toolchain-run.ps1 scripts/stream-crop-from-evidence.ts m2b-static-02
& ./scripts/toolchain-run.ps1 scripts/r34-differences.ts
& ./scripts/toolchain-run.ps1 scripts/make-hub-kit.ts
M3, from the same root (outputs packages/farm/dist/m3/{standalone,test,dev}, sizes in evidence/build/m3/; <new> is a new evidence label, since every runner refuses to overwrite):
& ./scripts/toolchain-run.ps1 scripts/build-farm.ts --release r34 --label m3 --all
& ./scripts/toolchain-run.ps1 scripts/check-farm-three-import.ts --label m3
& ./scripts/toolchain-run.ps1 scripts/check-bundle-hygiene.ts packages/farm/dist/m3/standalone
& ./scripts/toolchain-run.ps1 scripts/run-farm-mobile.ts --label <new> --backend both --root packages/farm/dist/m3/test
& ./scripts/toolchain-run.ps1 scripts/run-farm-tiering.ts --label <new> --backend both --root packages/farm/dist/m3/test
& ./scripts/toolchain-run.ps1 scripts/run-farm-play.ts --label <new> --backend both --root packages/farm/dist/m3/test --out-root evidence/m3/play
& ./scripts/toolchain-run.ps1 scripts/run-farm-contract.ts --release r34 --dist packages/farm/dist/m3 --out-root evidence/m3/contract --label <new> # add --webgl
& ./scripts/toolchain-run.ps1 scripts/make-device-kit.ts --label m3
pwsh -NoProfile -File packages/farm/dist/device-kit/check-device.ps1
pwsh -NoProfile -File packages/farm/dist/device-kit/run-tablet.ps1 -Steps state,b09,tier,soak,play,look [-Suffix <new>]
& ./scripts/toolchain-run.ps1 scripts/run-farm-look.ts --target pc --backend webgpu [--frame-tiers none] [--dest <dir>] # indicative only here
& ./scripts/toolchain-run.ps1 scripts/run-farm-look.ts --target tablet --backend webgpu --root packages/farm/dist/device-kit/scene/dev --frame-tiers none --dest evidence/tablet-<date>/look-webgpu-fullframe
& ./scripts/toolchain-run.ps1 scripts/run-farm-look.ts --target tablet --backend webgl2 --root packages/farm/dist/device-kit/scene/dev --capture-tiers none --dest evidence/tablet-<date>/look-webgl2
& ./scripts/toolchain-run.ps1 scripts/look-ab-views.ts
& ./scripts/toolchain-run.ps1 scripts/run-farm-look.ts --target pc --backend webgpu --presets default,neutral --views 'hero,house-porch,props' --frame-tiers none --capture-tiers high --all-view-tiers none --format png --dest evidence/look-ab/pc-webgpu
& ./scripts/toolchain-run.ps1 scripts/make-look-contact-sheet.ts --look-ab
& ./scripts/toolchain-run.ps1 scripts/run-farm-x04.ts --label <new> --runs 5 --cache cold # indicative only here; --trace records the trace
& ./scripts/toolchain-run.ps1 scripts/make-hub-kit.ts --label m3
& ./scripts/toolchain-run.ps1 run lint
& ./scripts/toolchain-run.ps1 scripts/update-m3-matrix.ts
& ./scripts/toolchain-run.ps1 scripts/render-report-matrix.ts
M4, from the same root (outputs packages/farm/dist/m4/{standalone,test,dev}, sizes in evidence/build/m4/; every timing here is indicative only):
& ./scripts/toolchain-run.ps1 scripts/build-farm.ts --release r34 --label m4 --all
& ./scripts/toolchain-run.ps1 scripts/run-farm-hitches.ts --label <new> --build packages/farm/dist/m4/test # cases webgl2:minimal,webgpu:minimal,webgl2:balanced; 3 revolutions; plain and traced runs
& ./scripts/toolchain-run.ps1 scripts/summarize-farm-hitches.ts --labels 'before,after,after2'
& ./scripts/toolchain-run.ps1 scripts/run-farm-instance-ab.ts --build packages/farm/dist/m4/test --dest evidence/m4/instance-ab
& ./scripts/toolchain-run.ps1 scripts/run-farm-x04.ts --label <new> --runs 3 --cache cold # then --cache warm; first stable frame, ready and declared
& ./scripts/toolchain-run.ps1 scripts/run-farm-x11-x12.ts --label m4 --alloc
& ./scripts/toolchain-run.ps1 scripts/check-farm-three-import.ts --label m4
& ./scripts/toolchain-run.ps1 scripts/check-bundle-hygiene.ts packages/farm/dist/m4/standalone
& ./scripts/toolchain-run.ps1 scripts/check-licenses.ts --bundle-modules packages/farm/dist/m4/standalone/bundle-modules.json
& ./scripts/toolchain-run.ps1 scripts/run-farm-contract.ts --release r34 --dist packages/farm/dist/m4 --out-root evidence/m4/contract --label <new> # add --webgl
& ./scripts/toolchain-run.ps1 scripts/run-farm-play.ts --label <new> --backend both --root packages/farm/dist/m4/test --out-root evidence/m4/play
& ./scripts/toolchain-run.ps1 scripts/run-farm-tiering.ts --label <new> --backend both --root packages/farm/dist/m4/test --out evidence/m4/tiering/<new>
& ./scripts/toolchain-run.ps1 scripts/run-farm-mobile.ts --label <new> --backend both --root packages/farm/dist/m4/test --out evidence/m4/b09/<new>
& ./scripts/toolchain-run.ps1 scripts/make-hub-kit.ts --label m4 --no-look
& ./scripts/toolchain-run.ps1 scripts/make-device-kit.ts --label m4
pwsh -NoProfile -File packages/farm/dist/device-kit/run-tablet.ps1 -Steps state,tier,soak -Date <date>
& ./scripts/toolchain-run.ps1 scripts/update-m4-matrix.ts
& ./scripts/toolchain-run.ps1 scripts/render-report-matrix.ts
The tablet commands need the tablet on USB in state device; they use the adb client only (never the adb server), map only the ports they use and remove them, and stop when the thermal status is not none. The device kit's README gives the method; results land in evidence/tablet-<date>/.
Quote comma lists in PowerShell (--views 'a,b'); unquoted, PowerShell joins them with spaces. Staging verifies the sealed archives, every listed member, the exact 23-model set and three-way GLB hashes; repeated no-force runs are identical, and r34 matches the website manifest's 23 model paths, byte counts and hashes (staging, release comparison). The sealed collider oracle for a release runs as $env:ORACLE_RELEASE='r34'; $env:ORACLE_COLLIDERS_ONLY='1'; & ./scripts/toolchain-run.ps1 packages/farm/scripts/audit-play-oracle.ts. Farm outputs are packages/farm/dist/{standalone,test,dev} (r33) and packages/farm/dist/r34/{standalone,test,dev}.
Acceptance matrix
Generated from the machine-readable matrix by scripts/render-report-matrix.ts. B-00 counts as eight subchecks and suffixed U ids count separately. The one B failure is the M0 B-00e defect with its required Z4 fallback. P-28 is page-owned. P-31 to P-35 are the M3 exit (SPEC 24) and pass. M4 (round 6) re-ran the counted checks on the m4 outputs and records X-11 and X-12; X-04 was redefined by the owner (D-23) and waits, with X-01 and X-05, for the coordinator's hub run of the m4 kit. Rows accepted under D-20 pass on the evidence gathered when the owner closed M2 as the record, and name the portion that evidence does not cover. Timing from the builder's PC is indicative only and decides no row.
| Family | Passing | Failing | Deferred |
|---|---|---|---|
| P | 33 | 0 | 2 |
| U | 32 | 0 | 0 |
| B | 21 | 1 | 1 |
| X | 7 | 0 | 5 |
| Id | Subject | State | Reason and evidence |
|---|---|---|---|
| P-01 | Boot and first complete frame | pass | B-01 passes on the Farm outputs of both releases (M2) and on the m3 outputs (M3), each on both backends (zero-size container, then onReady exactly once); every parity capture reached onReady with the six phase texts. Time to ready (X-04) was measured on the hub and fails on WebGPU (see X-04). M4 (round 6): B-01 passes on the m4 outputs on both profiles. Since M4 the scene compiles every pipeline of its first complete frame before onReady, behind the loading screen (SPEC 6.4 and 5.2(3) as amended by the owner on 2026-09-29 at 22:05, D-23): no program or pipeline is created after ready in any m4 hitch or X-04 run. X-04 now measures the first stable frame (D-23) and waits for the coordinator's hub run of the m4 kit (see X-04). r34-auto-m4/browser-results.json, r34-webgl2-m4/browser-results.json, hitches/README.md, hitches/attribution.json, m4/summary-cold.json, m4/summary-warm.json, r33-auto/browser-results.json, r33-webgl2/browser-results.json, r34-auto/browser-results.json, r34-webgl2/browser-results.json, r34-auto-m3-final/browser-results.json, r34-webgl2-m3-final/browser-results.json |
| P-02 | Manifest and verified loading | pass | U-01 and U-03 pass (kit); B-04 passes on the Farm outputs (404, hash mismatch, schema and CORS failures report the contract errors); both staged packs verify against the sealed receipts. m1/unit-coverage.md, r33-auto/browser-results.json, r33-webgl2/browser-results.json, r34-auto/browser-results.json, r34-webgl2/browser-results.json, staging/README.md, staging/r34.json |
| P-03 | Renderer and backend | pass | B-00a to c pass (M0); B-02 forced WebGL2 and B-03 WebGPU with adapter pass on the Farm outputs of both releases. m0/browser-results.json, r33-auto/browser-results.json, r33-webgl2/browser-results.json, r34-auto/browser-results.json, r34-webgl2/browser-results.json |
| P-04 | Lighting and environment | pass | B-06 static set: all 24 named-view/backend pairs pass at the unchanged thresholds with the Farm lighting (sun shadow camera set once). m2b-static-02/results.json, m2a-static-prepared-01/results.json, m2a-window-camera-02/results.json |
| P-05 | Texture pooling | pass | U-01b passes (kit); B-07 records 98 to 29 pooled textures and 39 GPU textures equal to the pilot at every static pair. m1/unit-coverage.md, m2b-static-02/results.json |
| P-06 | Placements and crop shadows | pass | U-19 (168 layout entries to 589 wrappers) passes; B-07 placements 589 and the crop shadow policy at every static pair. placements/README.md, m2b-static-02/results.json |
| P-07 | Clips and mixers | pass | U-19b initial clip table and offsets pass; B-07 at every static pair. placements/README.md, m2b-static-02/results.json |
| P-08 | Site mathematics | pass | Pure site mathematics matches seven sealed-pilot functions at 200 fixed points within 1e-6; U-05 is the complete acceptance requirement for this parity item world/summary.json |
| P-09 | Terrain and landscape | pass | U-06 terrain and landscape counts, groups and bounds match the oracle; B-06 static set passes. world/summary.json, m2b-static-02/results.json |
| P-10 | Woodland | pass | U-07 (843 trees, 16 cells) passes; B-07 8 packed groups, 3,372 instances, 64 source meshes; B-06 passes; the tangent-safe woodland A/B shows no material change (D-06). world/summary.json, m2b-static-02/results.json, m2a-static-prepared-01/woodland-ab-webgpu, m2a-static-prepared-01/woodland-ab-webgl2 |
| P-11 | Stream | pass | U-06 stream counts pass; B-06 passes, including the SPEC 19.7 stream crop at the mill bank (144 of 144 tiles on both backends); B-13 (M3) passes, including the stream crop after a governor step and a resize. world/summary.json, m2b-static-02/results.json, stream-crop/stream-crop.json, m3-final-webgpu/results.json, m3-final-webgl2/results.json |
| P-12 | Bridge | pass | U-06 bridge (34 boxes, 408 triangles) matches the oracle. world/summary.json |
| P-13 | Ground materials | pass | U-08 meadow pixels are byte-identical to the sealed recipe; B-06 static set passes. world/summary.json, m2b-static-02/results.json |
| P-14 | Field Grass | pass | B-07 14,000 tufts in 16 meshes at every static pair; B-06 passes. m2b-static-02/results.json |
| P-15 | Batching, freezing and frame graph | pass | U-13 passes (kit); B-07 batching, freeze and frame-graph goldens from the sealed r33 functions (83 groups, 54 dynamic, 2,000 batched sources; 555 placements and 4,191 nodes frozen; 575 hidden roots; R3-08 reconciles the SPEC counts) at every static pair. M4 (round 6): FARM-006 (a dynamic batch is rewritten and re-bounded only when a source moved), static usage for the Farm's dynamic batches and instance matrices read as instanced attributes (three's documented onNodeBuilderCreated hook) cut the programs on minimal from 178 to 26 and the buffer uploads from 924,752 to 336 bytes a frame; the look A/B against three's per-mesh uniform path passes (13 of 24 view and backend pairs pixel-identical, the other 11 within 80 pixels by one level, B-06 tile metric). instance-ab/result.json, hitches/README.md, hitches/attribution.json, KIT-REQUESTS.md, m4/unit-full.log, m1/unit-coverage.md, m2b-static-02/results.json |
| P-16 | Quality profiles | pass | The Farm tier table (tiers.ts, SPEC 15; live ladder grass only, R5-01) passes U-15 and the Farm tier tests; B-10 passes on m3 on both backends (emulated S24+, Tab S9 FE, WebGL2 phone, Save-Data and desktop profiles each start on the tier the table predicts); on the real tablet the classification from its own adapter strings assigns economy on WebGPU (arm, valhall) and minimal on WebGL2 (ANGLE Mali-G68, D-03), as predicted. M4 (round 6): X-07 (owner decision 2026-09-29 22:10, FARM-007): a phone or tablet on WebGPU whose adapter names a Mali-G5x or G6x, or names no Mali model with vendor arm and architecture valhall, starts on minimal, its WebGL2 tier; the kit's regression test and the Farm tier tests pass, B-10 on m4 passes on both backends with the Tab S9 FE profile now expecting minimal, and on the real tablet (m4, 2026-09-30) the classification from its own adapter strings now assigns minimal on WebGPU (vendor arm, architecture valhall; reason 'Mali-G68 class on WebGPU (X-07)') and minimal on WebGL2 (ANGLE Mali-G68, D-03), as predicted; the M3 result (economy on WebGPU) is retained for the old rule. m4-webgpu/results.json, m4-webgl2/results.json, tier/tier-auto.json, tier/tier-webgl2.json, quality/quality.test.ts, m4/unit-full.log, src/tiers.ts, m3-final-webgpu/results.json, m3-final-webgl2/results.json, m3-webgpu/results.json, m3-webgl2/results.json, tier/tier-auto-final.json, tier/tier-webgl2-final.json, tier/tier-auto.json, tier/tier-webgl2.json, m3/unit-full.log |
| P-17 | Visibility and teardown | pass | B-05 true teardown (C-04 protocol) and B-11 pause-when-hidden (second tab pauses) pass on the Farm outputs of both releases and both backends. M4 (round 6): B-05 teardown and B-11 pause-when-hidden pass on the m4 outputs on both profiles. r34-auto-m4/browser-results.json, r34-webgl2-m4/browser-results.json, r33-auto/browser-results.json, r33-webgl2/browser-results.json, r34-auto/browser-results.json, r34-webgl2/browser-results.json |
| P-18 | Orbit camera and views | pass | B-06: the 12 named views read from layout.views pass on both backends, including house-window-out with the interior polar rule (R4-01). m2b-static-02/results.json, m2a-static-prepared-01/results.json, m2a-window-camera-02/results.json |
| P-19 | Optional route playback | deferred | Dropped under the SPEC 24 cut line: D-09 keeps the tour out of the public scene, so its B-06 tour frames do not apply. The route itself is ported as the hub-kit living-route workload (test and dev builds only). play/workloads.ts, m2d/hub-kit-README.txt |
| P-20 | Collision world | pass | U-09 kit collision world and Farm selection pass; the 13 colliders are byte-identical to the sealed pilot (oracle); B-08 passes on both backends and its colliders equal the fixture. m1/unit-coverage.md, play/unit-full-b08.log, play/oracle.json, fixtures/play-colliders.json, b08-03-webgpu/results.json, b08-03-webgl2/results.json |
| P-21 | Walking | pass | U-10 passes (kit); 4,291 lockstep frames equal the sealed play.mjs; B-08 walk, run, strafe, wall, river and bridge steps pass on both backends. m1/unit-coverage.md, play/oracle.json, b08-03-webgpu/results.json, b08-03-webgl2/results.json |
| P-22 | Doors and gates | pass | U-11 (four mechanisms, damp 7, blocked-by-player revert, labels) passes; B-08 opens and closes every door mechanism on both backends. play/unit-full-b08.log, play/oracle.json, b08-03-webgpu/results.json, b08-03-webgl2/results.json |
| P-23 | Tractor and farmer rig | pass | U-14 (tractor kinematics, probes, exit order, joints; ground plan) passes; B-08 drive, steer, reverse, bound, obstacle, wedge and dismount pass on both backends. play/unit-full-b08.log, play/oracle.json, b08-03-webgpu/results.json, b08-03-webgl2/results.json |
| P-24 | Interaction prompt | pass | U-11 prompt radii and text table pass; B-08 prompt steps pass on both backends. play/unit-full-b08.log, b08-03-webgpu/results.json, b08-03-webgl2/results.json |
| P-25 | Play camera | pass | B-08 follow camera: drag turn without creep, wall obstruction pull-in to the oracle distance, subject hiding and recovery, on both backends. play/oracle.json, b08-03-webgpu/results.json, b08-03-webgl2/results.json |
| P-26 | Destinations and reset | pass | B-08: seven destinations and reset through the test hook and the HUD, refused while wedged; no public destination menu (D-09). b08-03-webgpu/results.json, b08-03-webgl2/results.json |
| P-27 | Herd motion | pass | U-22 herd positions and playback at 13 samples match the pilot fixture; B-07 at every static pair. placements/README.md, m2b-static-02/results.json |
| P-28 | Page-owned fullscreen | deferred | Page-owned by specification; no scene implementation or test required; listed deferred/out of scope preflight/stop-rule.md |
| P-29 | Camera buttons | pass | B-11 (44 px targets, contrast, keyboard reachability) passes on the Farm outputs; the overview CameraButtons render in the public HUD. r33-auto/browser-results.json, r33-webgl2/browser-results.json, r34-auto/browser-results.json, r34-webgl2/browser-results.json, ui/FarmHud.tsx |
| P-30 | Scoped keyboard input | pass | U-17 passes (kit); B-08 keyboard flows and B-11 root-scoped keys (Enter starts play, Escape twice, Tab leaves) pass on both backends. m1/unit-coverage.md, b08-03-webgpu/results.json, b08-03-webgl2/results.json, r33-auto/browser-results.json, r33-webgl2/browser-results.json, r34-auto/browser-results.json, r34-webgl2/browser-results.json |
| P-31 | New teardown | pass | B-05 (ten mount and unmount cycles, teardown and listener audit, C-03/C-04) passes on the m3 outputs on both profiles. M4 (round 6): B-05 passes on the m4 outputs on both profiles (heap growth after forced GC +1.684 % default-backend, +1.603 % forced WebGL2). r34-auto-m4/browser-results.json, r34-webgl2-m4/browser-results.json, r34-auto-m3-final/browser-results.json, r34-webgl2-m3-final/browser-results.json, r34-auto-m3/browser-results.json, r34-webgl2-m3/browser-results.json |
| P-32 | Mobile play | pass | B-09 passes on the final m3 build with Chrome touch emulation on both backends (16 of 16) and with real touch on the tablet (CDP Input.dispatchTouchEvent into the device Chrome, no emulation) on both backends (14 of 14; the safe-area and mouse-layout steps are emulation-only and skipped on the device). Earlier m3 runs are retained; there the first tablet WebGL2 attempt failed only because the WebGPU run's help dismissal persisted in the device Chrome's storage (the runner now clears its own origin first). D-22: joystick, drag look, pinch zoom, one context tap target, no button pads on touch. M4 (round 6): B-09 with touch emulation passes on the m4 outputs on both backends; the tablet flows were not repeated (M4 changed no input path). m4-webgpu/results.json, m4-webgl2/results.json, m3-final-webgpu/results.json, m3-final-webgl2/results.json, b09-webgpu-final/results.json, b09-webgl2-final/results.json, m3-webgpu/results.json, m3-webgl2/results.json, b09-webgpu/results.json, b09-webgl2/results.json, b09-webgl2-attempt2/results.json |
| P-33 | Adaptive tiering | pass | U-15 and U-16 pass; B-10 and B-13 pass on m3 on both backends; on the tablet the assigned tier holds with 0 governor changes and 0 locks over 210 s (X-10). M4 (round 6): U-15 and U-16 pass with the X-07 classification case and the three FARM-004 governor cases; B-10 and B-13 pass on m4 on both backends. On the tablet with the m4 build (2026-09-30) the assigned-tier soaks (minimal: 210 s of orbit, 60 s each of route and drive) make 0 changes, and forced high now steps down below 20 FPS: 3 changes in 210 s (0.86 a minute), 1 reversal and 1 lock, within X-10 (at most 4 changes a minute and one lock a session); in M3 the same forced-high soak made 0 changes. m4-webgpu/results.json, m4-webgl2/results.json, soak/assigned-living-orbit.json, soak/high-living-orbit.json, soak/assigned-living-route.json, soak/assigned-tractor-drive.json, quality/quality.test.ts, m4/unit-full.log, m3-final-webgpu/results.json, m3-final-webgl2/results.json, m3-webgpu/results.json, m3-webgl2/results.json, tier/tier-auto-final.json, tier/tier-webgl2-final.json, tier/tier-auto.json, tier/tier-webgl2.json, soak/assigned-living-orbit.json, m3/unit-full.log |
| P-34 | Public UI and accessibility | pass | B-11 and B-12 pass on the m3 outputs on both profiles; the touch HUD passes axe in overview, play, help, context and driving, 44 px targets and text contrast (B-09 P-34 step, emulated and on the tablet); the site's 390 px HUD overlap is fixed (B-09 layout matrix at 360 to 430 px, portrait and landscape, touch and mouse). M4 (round 6): B-11 and B-12 pass on the m4 outputs on both profiles, and the touch HUD checks of B-09 pass on m4 with emulation. r34-auto-m4/browser-results.json, r34-webgl2-m4/browser-results.json, m4-webgpu/results.json, m4-webgl2/results.json, m4/hygiene-m4.log, r34-auto-m3-final/browser-results.json, r34-webgl2-m3-final/browser-results.json, r34-auto-m3/browser-results.json, r34-webgl2-m3/browser-results.json, m3-final-webgpu/results.json, m3-final-webgl2/results.json, b09-webgpu-final/results.json, b09-webgl2-final/results.json |
| P-35 | Pause and reduced motion | pass | B-11 passes on the m3 outputs on both profiles: reduced motion freezes ambient motion, the page pauses when hidden (a second tab) and resumes, and the Farm door flow runs. M4 (round 6): B-11 passes on the m4 outputs on both profiles. r34-auto-m4/browser-results.json, r34-webgl2-m4/browser-results.json, r34-auto-m3-final/browser-results.json, r34-webgl2-m3-final/browser-results.json, r34-auto-m3/browser-results.json, r34-webgl2-m3/browser-results.json |
| U-01 | SHA-256 | pass | Executed deterministic M1 kit coverage; Farm-specific integration remains in its parity ids m1/unit-coverage.md |
| U-01b | Texture pooling | pass | Executed deterministic M1 kit coverage; Farm-specific integration remains in its parity ids m1/unit-coverage.md |
| U-02 | Asset URL resolution | pass | Executed deterministic M1 kit coverage; Farm-specific integration remains in its parity ids m1/unit-coverage.md |
| U-03 | Staging happy path | pass | Executed deterministic M1 kit coverage; Farm-specific integration remains in its parity ids m1/unit-coverage.md |
| U-04 | Staging failures | pass | Executed deterministic M1 kit coverage; Farm-specific integration remains in its parity ids m1/unit-coverage.md |
| U-05 | Site mathematics | pass | Seven site functions at 200 fixed points match the frozen and live sealed r33 oracle within 1e-6 world/summary.json |
| U-06 | Geometry generators | pass | Terrain, surrounding terrain, stream and bridge counts, groups, bounds, sampled attributes, component sums and exact integer index hashes match the sealed oracle world/summary.json |
| U-07 | Woodland | pass | All 843 woodland placements and 16 cell counts match the oracle; synthetic four-source-mesh packing yields 64 to 8 meshes with one shared tangent-safe derivative; loaded-Farm B-07 remains separate world/summary.json |
| U-08 | Meadow | pass | All 256px procedural meadow color, roughness and normal bytes match exact sealed-oracle hashes; material texture spaces, repeat and clone policy pass world/summary.json |
| U-09 | Collision world | pass | Kit collision world (M1) plus the Farm collider selection test; the 13 Farm colliders are byte-identical to the sealed pilot (r33 and r34). m1/unit-coverage.md, play/unit-full-b08.log, m2d/unit-full.log, play/oracle.json, fixtures/play-colliders-r34.json, m2d/colliders-r33.json, m2d/colliders-r34.json |
| U-10 | Capsule mover and fixed steps | pass | Executed deterministic M1 kit coverage; Farm-specific integration remains in its parity ids m1/unit-coverage.md |
| U-11 | Doors and interaction | pass | Four play.test.ts U-11 tests pass; the lockstep oracle exercises every door and prompt state. play/unit-full-b08.log, m2d/unit-full.log, play/oracle.json |
| U-12 | Tangent safety | pass | Executed deterministic M1 kit coverage; Farm-specific integration remains in its parity ids m1/unit-coverage.md |
| U-13 | Batching policy | pass | Executed deterministic M1 kit coverage; Farm-specific integration remains in its parity ids m1/unit-coverage.md |
| U-14 | Tractor | pass | Two play.test.ts U-14 tests pass (kinematics, probes, joints; ground plan and exit order); the lockstep oracle covers drive, wedge and dismount. play/unit-full-b08.log, m2d/unit-full.log, play/oracle.json |
| U-15 | Device classification and tiers | pass | Kit coverage (M1) and the Farm tier table tests (M3: S24+, Tab S9 FE, desktop, WebGL2 phone, Save-Data, the one-level tiers). M4 (round 6): the X-07 case (Tab S9 FE class on WebGPU starts on minimal; learned tiers never lift it; explicit props win; a table without minimal keeps the earlier result) passes in the kit's quality tests, with the Farm tier tests. quality/quality.test.ts, unit/tiers.test.ts, m4/unit-full.log, m1/unit-coverage.md, m3/unit-full.log |
| U-16 | Governor traces | pass | Kit coverage (M1); in the browser on m3, B-10 replays the governor traces through the Farm ladder. M4 (round 6): FARM-004: the governor evaluates once its window holds 60 samples or has spanned windowMs; the three new cases (a sub-20 FPS device steps down, reverts and locks, and advises a lower tier) failed on the 60-sample rule and pass; B-10 on m4 replays the traces. quality/quality.test.ts, m4-webgpu/results.json, m4-webgl2/results.json, m4/unit-full.log, m1/unit-coverage.md, m3-final-webgpu/results.json, m3-final-webgl2/results.json, m3/unit-full.log |
| U-17 | Input | pass | Executed deterministic M1 kit coverage; Farm-specific integration remains in its parity ids m1/unit-coverage.md |
| U-17b | Zone visibility | pass | Executed deterministic M1 kit coverage; Farm-specific integration remains in its parity ids m1/unit-coverage.md |
| U-18 | HUD store | pass | Kit HUD store (M1) plus the Farm string table test: pilot messages verbatim, four door labels, prompts, no coordinates. In M3 it caught a reworded touch tractor status (the D-22 HUD edit); the pilot message was restored (SPEC 13.1, INV 6) and the m3 outputs rebuilt. m1/unit-coverage.md, play/unit-full-b08.log, m2d/unit-full.log, m3/unit-full.log, m3/revert-tractor-touch-diff.json |
| U-19 | Layout expansion | pass | 168 entries expand to 589 wrappers with oracle-matching asset totals, positions, yaw, parentage, omitted paths and soil policy; source ownership and disposal cases pass placements/README.md |
| U-19b | Clip table | pass | Every initial clip/rest state and global-index time offset matches the sealed viewer; 22 actions, reversed watermill Spin, one-shot loops, action-only stepping and teardown pass placements/README.md |
| U-20 | Pin checking | pass | Executed deterministic M1 kit coverage; Farm-specific integration remains in its parity ids m1/unit-coverage.md |
| U-21 | Bundle hygiene | pass | Executed deterministic M1 kit coverage; Farm-specific integration remains in its parity ids m1/unit-coverage.md |
| U-22 | Herd | pass | All 20 animals match sealed-pilot poses and playback at 13 samples through 600 fixed steps within 1e-6; stop, restart, dt cap and failure cleanup cases pass placements/README.md |
| U-23 | Preset blender | pass | Executed deterministic M1 kit coverage; Farm-specific integration remains in its parity ids m1/unit-coverage.md |
| U-24 | Paths and orbit clamps | pass | Executed deterministic M1 kit coverage; Farm-specific integration remains in its parity ids m1/unit-coverage.md |
| U-25 | Chase camera | pass | Executed deterministic M1 kit coverage; Farm-specific integration remains in its parity ids m1/unit-coverage.md |
| U-26 | Scene clock | pass | Executed deterministic M1 kit coverage; Farm-specific integration remains in its parity ids m1/unit-coverage.md |
| U-27 | Developer parameters | pass | Executed deterministic M1 kit coverage; Farm-specific integration remains in its parity ids m1/unit-coverage.md |
| U-28 | Instancing core | pass | Executed deterministic M1 kit coverage; Farm-specific integration remains in its parity ids m1/unit-coverage.md |
| U-29 | Cell streamer | pass | Executed deterministic M1 kit coverage; Farm-specific integration remains in its parity ids m1/unit-coverage.md |
| B-00a | Async renderer and cube | pass | M0 correctness checks passed m0/browser-results.json |
| B-00b | Backend and adapter | pass | M0 correctness checks passed m0/browser-results.json |
| B-00c | Forced WebGL2 | pass | M0 correctness checks passed m0/browser-results.json |
| B-00d | Shadows prop | pass | M0 correctness checks passed m0/browser-results.json |
| B-00e | Pixel ratio and resize | fail | Confirmed viewport-size refresh defect on both backends; SPEC Z4 startup-only DPR fallback applied m0/browser-results.json |
| B-00f | StrictMode and late-init disposal | pass | M0 correctness checks passed m0/browser-results.json |
| B-00g | Disposal counters and loss | pass | M0 correctness checks passed m0/browser-results.json |
| B-00h | Neutral and AgX on both backends | pass | M0 correctness checks passed m0/browser-results.json |
| B-01 | Boot | pass | Kit demo (M1) and the Farm outputs of both releases on both backends, attempt 1. m1/qualification-round3.json, browser-auto-c03-c04/browser-results.json, browser-webgl2-c03-c04-callback/browser-results.json, r33-auto/browser-results.json, r33-webgl2/browser-results.json, r34-auto/browser-results.json, r34-webgl2/browser-results.json |
| B-02 | Forced WebGL2 | pass | Kit demo (M1) and the Farm outputs of both releases on both backends, attempt 1. m1/qualification-round3.json, browser-auto-c03-c04/browser-results.json, browser-webgl2-c03-c04-callback/browser-results.json, r33-auto/browser-results.json, r33-webgl2/browser-results.json, r34-auto/browser-results.json, r34-webgl2/browser-results.json |
| B-03 | WebGPU | pass | Kit demo (M1) and the Farm outputs of both releases on both backends, attempt 1. m1/qualification-round3.json, browser-auto-c03-c04/browser-results.json, browser-webgl2-c03-c04-callback/browser-results.json, r33-auto/browser-results.json, r33-webgl2/browser-results.json, r34-auto/browser-results.json, r34-webgl2/browser-results.json |
| B-04 | Error paths | pass | Kit demo (M1) and the Farm outputs of both releases on both backends, attempt 1, including the five injected callback failures. m1/qualification-round3.json, browser-auto-c03-c04/browser-results.json, browser-webgl2-c03-c04-callback/browser-results.json, r33-auto/browser-results.json, r33-webgl2/browser-results.json, r34-auto/browser-results.json, r34-webgl2/browser-results.json |
| B-05 | Teardown | pass | C-04 protocol on the Farm outputs, attempt 1: heap growth after forced GC r33 +1.79% (auto) and +1.67% (WebGL2), r34 +2.03% (auto) and +1.68% (WebGL2); one C-03 listener retained. M4 (round 6): m4 outputs, attempt 1: heap growth after forced GC +1.684 % (default backend) and +1.603 % (forced WebGL2); one C-03 listener retained. r34-auto-m4/browser-results.json, r34-webgl2-m4/browser-results.json, m1/qualification-round3.json, browser-auto-c03-c04/browser-results.json, browser-webgl2-c03-c04-callback/browser-results.json, r33-auto/browser-results.json, r33-webgl2/browser-results.json, r34-auto/browser-results.json, r34-webgl2/browser-results.json |
| B-06 | Parity screenshots | pass | Accepted under D-20 on the gathered set: 24 of 24 named-view/backend pairs (m2b-static-02, unchanged thresholds), play positions yard, house porch and bridge on WebGPU, stream crop on both backends, woodland A/B (M2a), and on r34 the farmhouse views and the farmhouse interior through the open door (play-house-door), 144 of 144 tiles on both backends against the sealed r34 pilot; r33 to r34 differences are the farmhouse floor only (D-07). M3: the WebGL2 play positions were captured on the tablet as screenshots (its minimal tier with the HUD, and the 1280 x 720 parity framing at tier high). Accepted under D-20. Not covered: The WebGL2 play positions (yard, house porch, bridge) are screenshots on the tablet, not compared with the pilot. The r33 play-house-door WebGPU capture failed 132 of 144 tiles because the focused root drew its outline in the rewrite frame; the capture now hides that outline (R4-07), it was not re-run on r33 under D-20, and the same view passes on r34 on both backends. m2b-static-02/results.json, m2a-static-prepared-01/results.json, m2a-window-camera-02/results.json, m2d-play-focused-02/results.json, stream-crop/stream-crop.json, m2a-static-prepared-01/woodland-ab-webgpu, m2a-static-prepared-01/woodland-ab-webgl2, m2d-r34-01/results.json, m2d-r34-01/r33-r34-differences.json, device/results.json, parity/results.json |
| B-07 | Statistics goldens | pass | Static goldens at all 24 static pairs; collider counts equal the frozen r33 fixture in B-08; r34: the sealed r34 collision world was frozen from the sealed r34 play.mjs (every soup byte-identical to the port; +264 static triangles, all in the r34 farmhouse) and the colliders of every r34 capture equal it, with all other r34 counts equal (R4-09; farmhouse-only differences, D-07). m2b-static-02/results.json, fixtures/play-colliders.json, fixtures/play-colliders-r34.json, m2d/colliders-r33.json, m2d/colliders-r34.json, b08-03-webgpu/results.json, b08-03-webgl2/results.json, m2d-r34-01/results.json, m2d-r34-01/r33-r34-differences.json |
| B-08 | Desktop play | pass | 24 of 24 steps on WebGPU and on forced WebGL2 with real keys and mouse: M2 (b08-03, r33) and M3 on the final m3 build (m3-final, r34 pack; its collider step uses the sealed r34 fixture, R4-09). Retained: the first m3 attempt passed every play step and failed only that step because the runner still read the r33 fixture; m3-02 then passed on the pre-revert m3 build. M4 (round 6): 24 of 24 steps on the m4 outputs on WebGPU and forced WebGL2, colliders equal to the sealed r34 fixture. m4-webgpu/results.json, m4-webgl2/results.json, fixtures/play-colliders-r34.json, b08-03-webgpu/results.json, b08-03-webgl2/results.json, m3-final-webgpu/results.json, m3-final-webgl2/results.json, m3-02-webgpu/results.json, m3-02-webgl2/results.json, m3-webgpu/results.json, m3-webgl2/results.json, play/oracle.json |
| B-09 | Mobile play | pass | Final m3 build. Emulated (412 x 915, DPR 2.8125, coarse pointer) on both backends: 16 of 16. Tablet with real touch on both backends: 14 of 14 plus two emulation-only steps skipped. The earlier m3 runs are retained (tablet WebGL2 attempt 1 failed only on the persisted help state; attempt 2 passed). M4 (round 6): m4 outputs with touch emulation on both backends: every step passes. m4-webgpu/results.json, m4-webgl2/results.json, m3-final-webgpu/results.json, m3-final-webgl2/results.json, b09-webgpu-final/results.json, b09-webgl2-final/results.json, m3-webgpu/results.json, m3-webgl2/results.json, b09-webgpu/results.json, b09-webgl2/results.json, b09-webgl2-attempt2/results.json |
| B-10 | Tiering | pass | Emulated profiles through the adapter-string hook start on the predicted tier; the improving trace gives exactly one down step and one up step after upHoldMs; a constant 30 ms trace steps down, reverts on the benefit rule and locks after two reversals; on both backends. M4 (round 6): m4 outputs on both backends; the Tab S9 FE profile now expects minimal on WebGPU (X-07). m4-webgpu/results.json, m4-webgl2/results.json, m3-final-webgpu/results.json, m3-final-webgl2/results.json, m3-webgpu/results.json, m3-webgl2/results.json |
| B-11 | Accessibility and lifecycle | pass | Kit demo (M1), the Farm outputs of both releases (M2) and the m3 outputs on both profiles: axe, focus and Enter starts play, the Farm walk and door flow, Escape twice, Tab leaves, aria-live, reduced motion, second tab pauses, 44 px targets, contrast. M4 (round 6): the m4 outputs on both profiles. r34-auto-m4/browser-results.json, r34-webgl2-m4/browser-results.json, m1/qualification-round3.json, browser-auto-c03-c04/browser-results.json, browser-webgl2-c03-c04-callback/browser-results.json, r33-auto/browser-results.json, r33-webgl2/browser-results.json, r34-auto/browser-results.json, r34-webgl2/browser-results.json, r34-auto-m3-final/browser-results.json, r34-webgl2-m3-final/browser-results.json, r34-auto-m3/browser-results.json, r34-webgl2-m3/browser-results.json |
| B-12 | Public and dev hygiene | pass | Kit demo (M1), the Farm outputs of both releases (M2) and the m3 outputs: the public build ignores dev and test flags and exposes no hooks (the look exploration is absent from public and test outputs); the dev build shows its panel. M4 (round 6): the m4 outputs on both profiles; the public m4 output passes the bundle hygiene check. r34-auto-m4/browser-results.json, r34-webgl2-m4/browser-results.json, m4/hygiene-m4.log, m1/qualification-round3.json, browser-auto-c03-c04/browser-results.json, browser-webgl2-c03-c04-callback/browser-results.json, r33-auto/browser-results.json, r33-webgl2/browser-results.json, r34-auto/browser-results.json, r34-webgl2/browser-results.json, r34-auto-m3-final/browser-results.json, r34-webgl2-m3-final/browser-results.json, r34-auto-m3/browser-results.json, r34-webgl2-m3/browser-results.json, m3/hygiene-m3.log |
| B-13 | Pixel ratio and resize | pass | A governor grass step plus a 960 x 540 to 1280 x 720 resize equals a fresh 1280 x 720 load with the same step: 144 of 144 tiles at eye-height and at the watermill, full frame and stream crop; a control (grass 1 against .75) fails the same gate (eye-height 120, watermill 139 of 144 tiles), so the metric sees the step; both backends. Z4: no pixel-ratio step. M4 (round 6): m4 outputs on both backends. m4-webgpu/results.json, m4-webgl2/results.json, m3-final-webgpu/results.json, m3-final-webgl2/results.json, m3-webgpu/results.json, m3-webgl2/results.json |
| B-14 | Site compatibility | deferred | Demo source consumer passes TypeScript 6.0.3 and one-three bundle assertion; Farm consumer remains required. M4 (round 6): the Farm consumer's check belongs to the site round (TASK-M4.md: the site's three alias and island check are not M4), so check:site-compat was not re-run (it re-checks the kit demo consumer and would overwrite this row's M1 evidence). The m4 public chunk built with the site's resolution (without the kit's three facade alias) is 1,934,965 B / 546,755 B gzip, 111.72 % / 108.02 % of the Farm ceiling, because three.module.js joins three/webgpu (+353,681 B / +84,334 B gzip); the Farm's own build has one three. m4/three-double-import.json, m1/site-compat.json |
| B-15 | Kit demo | pass | Complete M1 demo features and shared contract suite pass on WebGPU and forced WebGL2; Farm acceptance remains separate m1/qualification-round3.json |
| X-01 | Relative CPU frame time | pass | Hub run 2026-09-29 (the laptop, GTX 1660 Ti Max-Q, fullscreen 1920 x 1080 at 120 Hz, coordinator): all six backend and workload rows within 110 % of the pilot on frame-interval and CPU-render p50 and p95 (SPEC 20.1), measured on the r34 kit build. Not covered: the m4 build on the hub (the rebuilt kit, evidence/m4/hub-kit-README.txt; M4 adds the warm pass and the instanced-attribute path, so the coordinator's re-run of kit step 2 measures them). hub-2026-09-29/README.txt, results/comparison.json |
| X-02 | Rendering counts | pass | Draw calls within .38% and triangles within .024% of the pilot at every named view on both backends; pipelines equal to the pilot; play fixtures yard, house porch and bridge on WebGPU equal the pilot; on r34 the farmhouse views and play-house-door pass on both backends against the sealed r34 pilot. Accepted under D-20. Not covered: Play fixtures yard, house porch and bridge on forced WebGL2 not captured. m2b-static-02/results.json, m2d-play-focused-02/results.json, m2d-r34-01/results.json, m2d-r34-01/r33-r34-differences.json |
| X-03 | Leaks | pass | Farm B-05 ten-cycle heap growth below 5% on both releases and backends (see B-05). M4 (round 6): B-05 on the m4 outputs: +1.684 % (default backend), +1.603 % (forced WebGL2), under 5 %. r34-auto-m4/browser-results.json, r34-webgl2-m4/browser-results.json, m1/qualification-round3.json, browser-auto-c03-c04/browser-results.json, browser-webgl2-c03-c04-callback/browser-results.json, r33-auto/browser-results.json, r33-webgl2/browser-results.json, r34-auto/browser-results.json, r34-webgl2/browser-results.json |
| X-04 | Time to ready | deferred | Redefined by the owner on 2026-09-29 at 22:05 (D-23; SPEC 6.4 and 5.2(3) amended): the time to the first stable frame (the first frame from the ready frame on with no shader, program or pipeline creation, no asynchronous creation in flight and at most 50 ms) for both scenes, rewrite median within 120 % of the pilot's per backend and cache state, every WebGL2 rewrite run at most 45 s. The m4 rewrite compiles before ready (warm pass) and the hub kit and this PC's runner measure the new moment; the ready and declared moments stay as extra columns (x04Ready, x04Declared). The M3 hub result measured the old moment (WebGPU cold 1.316, warm 1.373; WebGL2 1.004 and 1.020) and is retained. On this PC (indicative only, 3 runs each, load 2 to 18 % CPU): first stable frame rewrite / pilot cold WebGPU 3,366 / 7,179 ms (0.469), cold WebGL2 4,158 / 19,453 ms (0.214), warm WebGPU 3,000 / 2,896 ms (1.036), warm WebGL2 3,445 / 3,466 ms (0.994); the pilot creates its pipelines in its ready frame. Deferred to the coordinator's hub run of kit steps 3 and 4. m4/summary-cold.json, m4/summary-warm.json, m4/hub-kit-README.txt, m4/hub-kit-MANIFEST.json, hub-2026-09-29/README.txt, results/comparison.json, x04-attempt/README.txt |
| X-05 | Desktop tier qualification | deferred | The m4 hub kit carries SPEC X-05 as its step 5: tiers high, balanced and economy, both backends, workloads living-orbit, living-route, walk and tractor-drive, 60 s of sampling after a 5 s warm-up, 5 runs; compare reports each case's median and range, SPEC 20.2's 60 FPS target per tier and the M4 proposed hitch rule (per 60 s run at every tier: 0 frames over 100 ms, 0 over 50 ms, longest at most 50 ms; not adopted). The M3 hub per-tier runs (30 s, 3 runs, no walk) are retained: no frame over 50 ms outside the minimal tier's first-sample compile stalls, which the M4 warm pass removes. Deferred to the coordinator's hub run and decision on the rule. m4/hub-kit-README.txt, m4/hub-kit-MANIFEST.json, hitches/README.md, hitches/attribution.json, hub-2026-09-29/README.txt, results/comparison.json, results/tiers |
| X-06 | Hub tier selection | deferred | Not covered by the hub kit: its runner forces the tier, so the automatic tier selection on the hub GPU is not exercised; it needs a run without a forced tier on the hub. m2d/hub-kit-README.txt, m2d/hub-kit-MANIFEST.json, m2d/hub-kit-smoke.json, m2d/hub-kit-smoke-attempt1.json, m2d/hub-kit-smoke-attempt1-diagnosis.txt, m2d/hub-kit-tier-mounts.txt |
| X-07 | Physical tablet | pass | Owner decision 2026-09-29 22:10 (D-25, FARM-007): the Tab S9 FE class starts on minimal on WebGPU, its WebGL2 tier, so SPEC 20.2's tablet target (30 FPS or better with p95 under 33.3 ms in orbit and in play, no task over 100 ms after ready) applies at minimal; SPEC 20.2's table still names economy as the expected tier. Measured on the owner's tablet with the m4 build at its assigned tier (minimal, WebGPU; the tablet's own frame intervals, valid with its device state: thermal status none, battery 93 to 92 % while charging, AP 34.5 to 39.1 °C): living-orbit 210 s at 59.2 FPS, p95 17.3 ms; living-route 60 s at 60.2 FPS, p95 17.5 ms; tractor-drive 60 s at 60.2 FPS, p95 17.2 ms; no task over 100 ms after ready and no frame over 100 ms; 2, 0 and 0 frames over 50 ms (longest 91.6, 33.9 and 31.3 ms). The M3 result at economy (orbit 30.2 FPS, p95 49.8 ms; failed) is retained. The pilot's tablet numbers (34.2 FPS orbit, 44.7 route, 83 ms worst gap) were recorded in an earlier session. Not covered: a same-session pilot run on the tablet (the pilot's numbers were recorded earlier). soak/assigned-living-orbit.json, soak/high-living-orbit.json, soak/assigned-living-route.json, soak/assigned-tractor-drive.json, tier/tier-auto.json, tier/tier-webgl2.json, soak/assigned-living-orbit.json, soak/high-living-orbit.json, soak/assigned-living-route.json, soak/assigned-tractor-drive.json |
| X-08 | Physical phone | deferred | No phone is connected. The device kit (packages/farm/dist/device-kit) drives any USB Android Chrome (KILN_TABLET_SERIAL selects the device), so the phone run is the tablet's commands on the owner's phone. device-kit/README.txt |
| X-09 | Soak | deferred | The 30-minute soak on the hub and phone is not part of the hub kit; deferred. PROGRESS.md |
| X-10 | Real-device governor stability | pass | FARM-004 is done (the governor evaluates on 60 samples or a spanned window). On the tablet with the m4 build (2026-09-30) the assigned-tier soaks (minimal: 210 s of orbit, 60 s each of route and drive) make 0 changes, and forced high now steps down below 20 FPS: 3 changes in 210 s (0.86 a minute), 1 reversal and 1 lock, within X-10 (at most 4 changes a minute and one lock a session); in M3 the same forced-high soak made 0 changes. Not covered: the phone (no device). soak/assigned-living-orbit.json, soak/high-living-orbit.json, soak/assigned-living-route.json, soak/assigned-tractor-drive.json, soak/assigned-living-orbit.json, soak/high-living-orbit.json, soak/assigned-living-route.json, soak/assigned-tractor-drive.json, KIT-REQUESTS.md |
| X-11 | Bytes | pass | Recorded on the m4 outputs (bytes only, no timing; headless 1920 × 1080, Chrome 154): the staged r34 pack is 34 files of 7,326,021 B and its 23 GLBs 7,098,444 B (SPEC's 7,070,552 B is the r33 pack's, which stays staged); the public m4 chunk is 1,581,284 B / 462,421 B gzip, 91.30 % / 91.36 % of the Farm's D-15 ceiling (1,732,040 B / 506,143 B). Before onReady the public page fetches 29 requests with 8,906,023 B of bodies (page 1,858, code 1,581,284, JSON 56,437, GLBs 7,098,444, grass binary 168,000; 8,914,723 B with headers) and nothing after onReady; the test build fetches the same 29 requests (8,918,214 B; its chunk is 12,191 B larger) at every tier on both backends, and nothing after onReady. In the site's resolution without the kit's three facade alias the chunk would be 1,934,965 B / 546,755 B (111.72 % / 108.02 % of the ceiling); the remedy is the site's three alias, which TASK-M4 leaves to the site round. m4/x11-bytes.json, m4/bundle-r34-public.json, m4/bundle-r34-test.json, m4/bundle-r34-dev.json, m4/three-double-import.json, staging/README.md |
| X-12 | Allocation trend | pass | 1,000 frames each of living-orbit and walking on the m4 test build (WebGPU and forced WebGL2, tiers high and minimal, headless 1920 × 1080), a forced collection (CDP HeapProfiler.collectGarbage) before and after, Runtime.getHeapUsage: used-heap growth 0.074 to 0.172 MB in all 8 cases (WebGPU high orbit 0.170 and walk 0.097, minimal 0.151 and 0.091; WebGL2 high 0.172 and 0.074, minimal 0.111 and 0.080), each with the camera moving, Rowan walking on foot and the tier held; limit 2 MB. Where the per-frame allocation is made (HeapProfiler sampling, collected objects included) is recorded beside it as a diagnostic, not a rule. m4/x12-allocation.json, m4/x12-allocation-sites.json |
Validation, parity thresholds and limits
Units, types and lint (M4, 2026-09-30 04:35 to 04:38 UTC; gate log). The full CPU suite (bun test, the whole workspace including the other builders' packages as they stood then) passes 343 tests, 1 optional oracle case skipped, 0 failures, 3,634,299 assertions across 58 files (log); this builder's packages alone pass 236, 1 skipped, 0 failures, 66,274 assertions across 40 files, including the FARM-004, FARM-006 and X-07 cases and the warm-pass and instanced-attribute tests (log). bun run typecheck exits 0 with no diagnostic (log); bun run lint reports 0 findings, gated and advisory (log). The public m4 output passes the bundle hygiene check and the production licence audit (see Bundle sizes). The final typecheck and lint after the last script edits are in the command log.
Units, types and lint (M3 close, 01:01 UTC). The full CPU suite (bun test, the whole workspace including the other builders' packages) passes 311 tests, 1 optional oracle case skipped, 0 failures, 1,764,849 assertions across 53 files (log); this builder's packages alone (bun test packages/farm packages/scene-kit scripts) pass 218, 1 skipped, 0 failures, including the FARM-005 regression scripts/tests/static-server.test.ts (log). bun run typecheck (TS7, tsc --noEmit -p tsconfig.json, workspace-wide) exits 0 (log). bun run lint (R5-06) reports 0 gated findings in packages/scene-kit, packages/farm and scripts, and one advisory unused-name finding in packages/golden-gate, which belongs to another builder and is not touched here (log). The same gates passed at 00:24 UTC, before FARM-005 and the frame-series change (270 tests across 46 files; units, typecheck, lint). The suite caught one M3 regression, U-18 (the D-22 HUD edit had reworded the pilot's touch tractor status); it is fixed (R5-11). At the M2 close: 199 tests / 65,568 assertions / 35 files (log). At M1 release: 132 tests / 6,215 assertions / 19 files; TS7 and the TS6 source consumer pass; all 125 explicit export names resolve (M1 units, exports).
Parity rule (unchanged since M2a). Linear luminance on 16 × 9 tiles; a tile passes within max(3 × old-old noise, .02); a pair passes with at least 97% of tiles and a global mean below .02. Captures are 1280 × 720 at DPR 1, tier high, governor frozen, herd off in the rewrite; the sealed pilot is served by its unmodified serve.mjs; each capture is 3 s after readiness with a repeat 3 s later for the pilot's old-old noise. Since R4-07 (stage m2d) the rewrite takes the pilot's shader-clock phase, bracketed around the pilot capture, and the capture CSS hides focus outlines. No threshold was relaxed or tightened.
| Set | Result | Evidence |
|---|---|---|
| M2a static, 12 views × 2 backends | 24/24 pass, 144/144 tiles each; house-window-out after the explained minimum-distance correction; woodland A/B: 0 tangent offenders safe versus 16 unsafe, no material visual change | static, window, woodland WebGPU |
| M2b static, optimized | 24/24 pass (141–144 tiles, global means at most .0107); X-02 draw calls within .38%, triangles within .024%, pipelines equal to the pilot; B-07 goldens at every pair (589 placements; 843 trees in 16 cells, 8 packed groups, 3,372 instances; 14,000 tufts in 16 meshes; 83 groups, 54 dynamic, 2,000 batched sources; 575 hidden roots; textures 98 to 29, 39 GPU textures). props/WebGPU failed attempt 1 on wind-animated grass (139/144) and passed the SPEC 18 retry (141/144); both retained | set |
| M2c play | Lockstep 4,291 frames with zero difference from the sealed play.mjs; 13 colliders byte-identical; B-08 24/24 steps on WebGPU and forced WebGL2 with real keys and mouse; b08-01/b08-02 retained |
oracle, WebGPU, WebGL2 |
| M2d play positions, WebGPU | yard 143/144 (global .0067, noise .0039), house porch 144/144 (.0032 / .0008), bridge 144/144 (.0106 / .0075), X-02 equal to the pilot. The r33 door view failed 132/144 twice because the focused root drew its outline (retained; normalized in R4-07) | set |
| M2d stream crop at the mill bank (SPEC 19.7) | 144/144 on both backends (WebGPU .0071 against noise .0106; WebGL2 .0089 against .0090), from the retained M2b captures | crop |
| M2d r34 staging run (SPEC 19.8) | house-interior, house-porch, house-window-out and play-house-door: attempt 2 of every pair passes 144/144 tiles and X-02 on both backends (global means .00006–.00069 against noise .00019–.00085), and B-07 against the sealed r34 collision world on the offline re-evaluation (R4-09). Both attempts are stored as failures because the run compared colliders with the r33 fixture. r33 to r34: the farmhouse floor only | run, differences and re-evaluation |
| Contract suite (B-01–B-05, B-11, B-12) | pass on attempt 1 for r33 and r34, default-backend and forced-WebGL2 profiles (Chrome 154.0.8037.58) | r33 default, r33 WebGL2, r34 default, r34 WebGL2 |
M3 B-09 mobile play, final m3 build |
Emulated touch (412 × 915, DPR 2.8125, coarse pointer): 16/16 steps on WebGPU and on forced WebGL2. Tablet with real touch (CDP Input.dispatchTouchEvent into the device's Chrome at its own screen size, no emulation): 14/14 on both backends, the two emulation-only steps (safe-area override, mouse layouts) skipped. Steps: touch profile; overview with no camera pad and one Reset view target; start play by tap, help once; joystick walk, run and strafe; drag look; pinch 1.6 to 10 m; the context button on the farmhouse door; tractor by touch; the mode button leaves play; touch HUD accessibility (axe, 44 px, contrast, live status); overview pinch and two-finger pan; HUD layout at 360 to 430 px portrait and landscape; safe areas; page scroll outside the scene; mouse layouts at 360 to 430 px; no browser errors. The earlier m3 runs are retained, including tablet WebGL2 attempt 1 (failed only on the help state the WebGPU run had persisted) |
emulated WebGPU, emulated WebGL2, tablet WebGPU, tablet WebGL2 |
M3 B-10 and B-13, final m3 build |
5/5 on both backends: emulated S24+, Tab S9 FE, WebGL2 phone, Save-Data and desktop profiles start on the tier the table predicts; an improving trace gives one down step, then one up step only after upHoldMs; a constant 30 ms trace steps down, reverts on the benefit rule and locks after two reversals (C-02); B-13: a governor grass step plus a 960 × 540 → 1280 × 720 resize equals a fresh 1280 × 720 load with the same step, 144/144 tiles at eye-height and at the watermill (full frame and stream crop); the control (grass 1 against .75) fails the same gate (120 and 139 of 144 tiles), so the gate sees the step. Z4: no pixel-ratio step |
WebGPU, WebGL2 |
M3 B-08, final m3 build |
24/24 steps on WebGPU and forced WebGL2; colliders equal the sealed r34 fixture (R4-09). Retained: m3 attempt 1 (the runner read the r33 fixture for an r34 build) and m3-02 (passed on the pre-revert build) |
WebGPU, WebGL2 |
M3 contract suite (B-01–B-05, B-11, B-12, Farm door flow), final m3 build |
pass on attempt 1, default-backend and forced-WebGL2 profiles (Chrome 154.0.8037.58); the pre-revert m3 runs also passed |
default, WebGL2, log |
| M3 WebGL2 play positions, tablet | yard, house porch and bridge on forced WebGL2 as screenshots: the tablet's own view at its minimal tier with the HUD, and the 1280 × 720 parity framing at tier high. Not compared with the pilot | device view, parity framing |
M4 contract suite (B-01–B-05, B-11, B-12, Farm door flow), m4 build |
pass on attempt 1, default-backend and forced-WebGL2 profiles (Chrome 154.0.8037.58); B-05 +1.684 % and +1.603 % | default, WebGL2, log |
M4 B-08, m4 build |
24/24 steps on WebGPU and forced WebGL2; colliders equal the sealed r34 fixture | WebGPU, WebGL2 |
M4 B-10 and B-13, m4 build |
5/5 on both backends; the Tab S9 FE profile now predicts minimal on WebGPU (X-07, FARM-007) | WebGPU, WebGL2 |
M4 B-09 with touch emulation, m4 build |
16/16 steps on WebGPU and forced WebGL2; the tablet's touch flows were not repeated (M4 changed no input path) | WebGPU, WebGL2 |
| M4 instanced-attribute look A/B (R6-03) | 12 sealed views × 2 backends at high, 1280 × 720, attributes against three's per-mesh uniform path: 13 of 24 pairs pixel-identical, the other 11 within 80 pixels by one level, all passing the B-06 tile metric | result |
Not covered, accepted under D-20: a pixel comparison of the WebGL2 play positions with the pilot (M3 captured them on the tablet as screenshots), and a rerun of the r33 door view (the same view passes on r34 on both backends). All failed attempts remain in their evidence directories.
Heap findings, timing and hub kit
Heap (memory, not timing). M1 found and fixed a real r186 leak (the shared DFG_LUT texture retained retired renderer listeners); cold raw heap growth from compiled code remains a historical failure under its original protocol. C-04 endorses R2-13's fixed initialization (120 frames per backend, ten initialization cycles, forced GC, ten measured alternating cycles plus loading/error cancellation, a third ten-cycle block for trend only). Retainers, fixed diagnostic, M1 qualification. The kit demo passes at +1.761% and +1.592%. The Farm passes on every release and profile, each with exactly the one C-03 listener, zero resources, zero Fiber roots and the owned device or context lost:
| Farm B-05 | Measured baseline B | End of measured block B | Growth B / percent | Third-block growth B | Investigate before M4 |
|---|---|---|---|---|---|
| r33, default-backend profile | 12,035,660 | 12,251,168 | 215,508 / 1.791% | 180,880 | No |
| r33, forced-WebGL2 profile | 12,046,584 | 12,247,480 | 200,896 / 1.668% | 179,528 | No |
| r34, default-backend profile | 12,063,668 | 12,308,820 | 245,152 / 2.032% | 152,000 | No |
| r34, forced-WebGL2 profile | 12,083,472 | 12,286,128 | 202,656 / 1.677% | 184,176 | No |
| m3 before the U-18 revert, default-backend profile | 12,083,676 | 12,323,708 | 240,032 / 1.986% | 163,616 | No |
| m3 before the U-18 revert, forced-WebGL2 profile | 12,074,780 | 12,301,496 | 226,716 / 1.878% | 181,948 | No |
| m3 final, default-backend profile | 12,087,472 | 12,308,780 | 221,308 / 1.831% | 181,000 | No |
| m3 final, forced-WebGL2 profile | 12,098,968 | 12,303,960 | 204,992 / 1.694% | 168,136 | No |
| m4, default-backend profile | 12,932,580 | 13,150,364 | 217,784 / 1.684% | 121,256 | No |
| m4, forced-WebGL2 profile | 12,920,736 | 13,127,908 | 207,172 / 1.603% | 144,540 | No |
The m4 baseline after the ten initialization cycles is about 0.83 MB above m3's on both profiles while the growth over the measured cycles stays where it was, so the extra heap is retained once per page (not per mount); it was not traced further in M4 (a candidate is module-level state that the warm pass fills by compiling every material before ready).
Timing. No timing from this PC settles anything. The hub numbers are the coordinator's run of the r34 kit (2026-09-29, 17:25 to 19:08 UTC, the owner's laptop hub: GTX 1660 Ti Max-Q, Chrome 150 headed fullscreen at 1920 × 1080 and 120.11 Hz, the pilot's quiet-gate preflight before every command group; README, comparison). X-01 passes on both backends: every backend × workload row is within 110 % of the pilot on frame-interval and CPU-render p50 and p95 (frame p50 8.3 ms, the refresh interval, on both sides). X-04 passes on WebGL2 and fails on WebGPU: WebGL2 cold 6314 against 6289 ms (1.004) and warm 4955 against 4857 ms (1.020), worst rewrite run 6414 ms against the 45 s limit; WebGPU cold 6013 against 4569 ms (1.316) and warm 5875 against 4279 ms (1.373) for a 1.20 limit. The per-tier step (72 files) gives the rewrite a frame p50 / p95 of 8.3 / 8.4 ms on WebGPU at every tier and workload, and on WebGL2 a p95 of 16.6–16.7 ms for living-orbit at high, balanced and economy and for living-route at high and balanced, 8.4 ms elsewhere; SPEC X-05 asks for 60 s windows, 5 runs and a walk workload, so X-05 stays deferred for the coordinator. These runs measured the r34 kit build (the M2 scene plus the hub workloads); neither the m3 nor the m4 build has been on the hub. The X-04 result above measured the moment the owner replaced at 22:05 (D-23; see the M4 section for the first stable frame and the m4 kit). The tablet's own numbers are in the Tablet sections; the look options' PC frame time and the X-04 attempt's PC timings are labelled indicative with their load samples.
X-04 attempt (TASK-M3 item 7; 18:25 to 18:58 EDT, about 35 minutes of the two-hour bound; notes, cold, warm). Every number in it is indicative only: this PC is shared (RTX 3070 on D3D12; another builder was working), and each run file carries its load sample. scripts/run-farm-x04.ts served the interim m3 test build and the sealed r34 pilot from the hub kit on its own 44xx server, drove headless Chrome 154 at 1920 × 1080 for five cold (fresh profile) and five warm runs per scene and backend, and recorded two readiness moments: next-frame, the first animation frame at which readiness is visible (the hub kit's X-04 measure), and declared, the moment the scene declares readiness (the rewrite's harness clears its status line inside onReady; the pilot writes its ready line inside the frame callback of its first render). One extra cold run per scene and backend recorded a Chrome trace. The dominant cost is the per-pipeline shader build (R-02), the same in both scenes: both create the same 234 render pipelines while submitting their first complete frame; three's TSL-to-WGSL node build is the largest main-thread cost (2.4 s rewrite, 2.6 s pilot), and Dawn then compiles each new pipeline's shaders with DXC, one after another (229 compiles, 6.5 s here). Pack fetch and GLB decode, texture upload and React commit are each well under a second. Why the WebGPU ratio fails: SPEC 5.2(3) fires onReady from a requestAnimationFrame scheduled inside the first complete frame, and the runner sees it one frame later, after the GPU compile stall; the pilot writes its ready line inside its first complete frame, before that stall. Both scenes show their first picture only after the stall. Medians here, pilot / rewrite / ratio: cold WebGPU next-frame 3429 / 7314 ms / 2.133 but declared 3430 / 3383 ms / 0.986; cold WebGL2 0.948 on both measures (no stall between frames: program compile and link block inside the first render); warm 0.96 to 0.97 on both backends (Dawn's shader cache removes the DXC work here; on the hub the warm ratio stayed at 1.373, so the hub's stall persists warm there, which cannot be verified from here). The one change is look-neutral and in the measurement (R5-12): the hub runner records declaredReadyMs beside readyMs in every X-04 file, and compare writes x04Declared beside x04 with the same median and 120 % rule; everything else is unchanged, so the 2026-09-29 hub files stay comparable. Not made, each an owner decision: asynchronous pipeline pre-compilation before onReady (compileAsync; SPEC 6.4 forbids prewarm, INV 4.3), onReady inside the first complete frame (SPEC 5.2(3), and a kit contract change for Golden Gate too), fewer distinct pipelines (three.js internals, a risk to the look). The attempt stops there: hub kit steps 3 and 4 re-run with the m3 kit give x04Declared on the hub, and the owner decides which moment X-04 measures (SPEC 20 names it "time to onReady").
Hub run kit, rebuilt for M3 (scripts/make-hub-kit.ts --label m3, packages/farm/dist/hub-kit/, 203 files, 35,570,394 bytes after the FARM-005 rebuild at 20:57 EDT, every file hashed in MANIFEST.json; README, manifest, build log): the sealed r34 scene archive written member by member (served root pilot/scene), the rewrite's final m3 test build (rewrite/, index-y-h2nCPw.js) and dev build (rewrite-dev/, with the look toggles), runner/perf-farm.mjs (README steps 0 to 5 as before: verify, smoke, X-01 as 36 files, X-04 as 40 files, compare, optional per-tier runs) and runner/look-farm.mjs (step 6: the look options' frame time per tier and the contact-sheet captures on the laptop, TASK-M3 item 6). The X-04 runs now also record the declared readiness (declaredReadyMs), and compare writes x04Declared beside x04 (R5-12). Every rewrite frames file also keeps its raw interval series (frameIntervalsMs) and framesOver100Ms beside p50, p95, p99, max, framesOver50Ms and longTasks, and the look runner keeps each cell's intervals (R5-15, the owner's note of 20:35); the sealed pilot's page reports only its summary, so both are null there. Deviations are unchanged from M2: 30 s windows because the sealed pilot harness samples 30 s, no walk workload because the pilot has no sealed walk measurement, and the rewrite plays its initial clip table during moving workloads. Pass rules are SPEC 20.3's, quoted in the README; results are copied back unchanged to evidence/perf/hub-<date>/.
Local plumbing checks on this PC, headless, every number discarded. On the first m3 build (20:22): smoke attempt 1 failed only rewrite living-orbit, because the workload was not yet running when sampling began (workloadRunningAtSamplingStart: false) while the contract suite was loading this PC, which makes the run invalid by the kit's own rule, so it was repeated, not shortened; attempt 2 passed all ten checks (attempt 1, attempt 2). A short look-runner check (two presets at one tier, one capture) completed its run but printed a server address, left a second server listening on 4400 and never exited (FARM-005: the bundled scripts/static-server.mjs started its command-line server inside runner/look-farm.mjs); the builder stopped its own process, fixed the guard with a regression test (R5-14) and rebuilt the kit: against the first build's manifest only README.txt, runner/perf-farm.mjs and runner/look-farm.mjs changed. On the rebuilt kit the smoke passes 10/10, including the kept interval series (smoke); the look-runner check completes and exits by itself with no listener left (look-runner checks); verify finds 203 files and no problems (verify). The M2 kit's record stays in its README, smoke, attempt 1 diagnosis and tier mounts.
M4 (round 6): WebGL2 hitches, warm pass, governor floor, bytes and allocation
Every timing in this section comes from this PC (shared, under load) or the owner's tablet and is indicative only; each run file carries its load sample, and nothing here settles a budget. The hub runs stay the coordinator's.
WebGL2 hitches on minimal (TASK-M4 item 1; method, finding, tables and rule, attribution table). scripts/run-farm-hitches.ts serves a build on an owned port and drives headless Chrome 154 at 1920 × 1080, DPR 1 (rAF at 144 Hz here) through three living-orbit revolutions (35 s each) from the first frame after ready, with the tier forced and the governor live. Per frame it keeps the rAF interval, the JS time, WebGL and WebGPU call counts and bytes (uploads, texture and render-target allocation, shader compiles, program links, pipeline and shader-module creation, synchronous GL or GPU call time), governor events and the scene step, and the heap over CDP; a second run per case records a Chrome trace (V8 GC, gpu.angle, gpu.dawn, timeline) from 30 s into revolution 1, and every frame over 50 ms is attributed from it. Cases: WebGL2 minimal, and for contrast WebGPU minimal and WebGL2 balanced; a plain and a traced run each.
- Finding. After the first revolution no frame exceeded 50 ms in any run, before or after the fix. The owner's hitches are in the first revolution: on WebGL2 minimal 6 frames over 100 ms (243.1 ms at worst) in two clusters of three, at the start of the orbit and 8.2 s in (about 84°), each linking one or two programs at about 100 ms of synchronous GL time per link, when the first woodland cells and batches of the far side come into view. WebGPU minimal creates the same 8 pipelines asynchronously (27.8 ms at worst); WebGL2 balanced had none. The hub's M3 per-tier runs show the same in their first sample (WebGL2 minimal orbit run 1: 3 frames over 50 ms, 216.6 ms; WebGPU minimal orbit: one frame of 58.5 to 75.1 ms per run; nothing else over 50 ms). Texture or mipmap uploads, shadow updates, garbage collection, governor changes, ANGLE recompiles, render-target reallocation and the 250 ms clamp produced no long frame; the link frames lost 393 to 455 ms of scene time to the kit's 100 ms step clamp.
- Cause. three 0.186.0's per-mesh instance uniform buffers (see the three evidence section): 178 programs on minimal and 211 on balanced, each compiled when its mesh first comes into view, and 924,752 bytes of uniform uploads a frame on minimal.
- Fix, look-neutral and within SPEC (R6-01, R6-03, R6-04, R6-07). The SPEC 6.4 warm pass before ready, bounded to 8 chains (with all 221 chains at once it left about 250 MB of node-builder garbage, whose one major collection made 20.9 and 27.7 ms frames in revolution 3; concurrency A/B: 8 was at or near the shortest time to ready in every case, with identical programs and pipelines); instance matrices as instanced attributes (26 programs on minimal, 40 on balanced); static usage for the Farm's dynamic batches with FARM-006. The look A/B against three's uniform path passes (result: 12 sealed views × 2 backends at high, 1280 × 720; 13 of 24 pairs pixel-identical, the other 11 within 80 pixels by one level, passing the B-06 tile metric). D-18 is untouched: MSAA 4x and ACES .95 stay the default on every tier.
| Case, three revolutions, before (m3) → after (final m4) | Frames over 50 / over 100 ms | Longest frame, plain / traced | Programs at start | Buffer uploads a frame | Heap at orbit start (plain) | Load before the after runs, plain / traced (CPU, GPU) |
|---|---|---|---|---|---|---|
| WebGL2 minimal | 6 / 6 → 0 / 0 | 243.1 / 236.1 → 7.9 / 13.9 ms | 178 → 26 | 924,752 → 336 B | 71 → 43 MB | 28.6 % 10 %, 17.9 % 1 % |
| WebGPU minimal | 0 / 0 → 0 / 0 | 27.8 / 27.8 → 7.6 / 13.9 ms | 178 → 26 | 924,752 → 336 B | 157 → 66 MB | 14.1 % 9 %, 24.9 % 15 % |
| WebGL2 balanced | 0 / 0 → 0 / 0 | 20.9 / 14.1 → 21.3 / 20.9 ms | 211 → 40 | 1,066,196 → 40,172 B | 42 → 44 MB | 17.5 % 2 %, 13.7 % 2 % |
After the fix no program or pipeline is created after ready, no long task runs, no governor level changes and no scene time is lost to the step clamp in any run; synchronous GL time falls from 0.39 to 0.24 ms a frame on WebGL2 minimal and GPU call time from 0.62 to 0.07 ms on WebGPU minimal. WebGL2 balanced keeps its occasional three-vsync frames (about 21 ms at 144 Hz) with no collection or creation beside them. The README keeps the intermediate after build (all chains at once) and the load samples of the before runs (9 to 30 % CPU before each run).
Proposed hitch rule (X-05 shape, R6-08; for the coordinator to adopt or not). Per 60 s run of the hub kit's step 5, the same at every tier and counted from the first sampled frame: 0 frames over 100 ms, 0 frames over 50 ms, longest frame at most 50 ms (PROPOSED_HITCH_RULE in packages/farm/scripts/perf.ts; compare reports its verdict per case). No first-revolution allowance is needed since no creation happens after ready; here the longest frame after the fix was 21.3 ms and the longest major collection inside an orbit 8.0 ms; on the hub every M3 run without a compile stall stayed at or under 33.4 ms, while the stalls this fix removes were 58.5 to 216.6 ms. If the first hub run shows frames over 50 ms from outside the scene, the over-50 count per minute is the one limit to relax. A mobile row needs device calibration (the tablet below gives the first numbers).
X-04 as the time to the first stable frame (TASK-M4 item 1b; owner decision 22:05, D-23; R6-02). SPEC 6.4 and 5.2(3) now say the scene compiles every pipeline for its first complete frame before ready, and X-04 measures the first frame from the ready frame on with no creation, nothing asynchronous in flight and at most 50 ms, for the pilot and the rewrite alike; the ready and declared moments stay as extra columns. Local indicative runs (scripts/run-farm-x04.ts --label m4 --runs 3, m4 test build against the sealed r34 pilot from the hub kit, headless 1920 × 1080, load 2 to 18 % CPU; cold, warm, log), median rewrite / pilot:
| Backend, cache | First stable frame (X-04 since D-23) | Next-frame ready (extra) | Declared ready (extra) |
|---|---|---|---|
| WebGPU, cold | 3,366 / 7,179 ms (0.469) | 3,359 / 3,201 ms (1.049) | 3,362 / 3,199 ms (1.051) |
| WebGL2, cold | 4,158 / 19,453 ms (0.214; worst rewrite run 4.2 s) | 4,123 / 19,440 ms (0.212) | 4,127 / 19,439 ms (0.212) |
| WebGPU, warm | 3,000 / 2,896 ms (1.036) | 2,993 / 2,876 ms (1.041) | 2,999 / 2,877 ms (1.042) |
| WebGL2, warm | 3,445 / 3,466 ms (0.994) | 3,438 / 3,445 ms (0.998) | 3,440 / 3,448 ms (0.998) |
The rewrite's first stable frame is its ready frame in every run; the pilot creates its pipelines in its ready frame, so on a cold WebGPU load its first stable frame follows 4 s later. The hub re-run (kit steps 3 and 4, then compare) is the coordinator's; X-04 is deferred to it.
Hub run kit, rebuilt for M4 (scripts/make-hub-kit.ts --label m4 --no-look, packages/farm/dist/hub-kit/, 160 files, 23,564,125 bytes, every file hashed in MANIFEST.json; README, manifest, build log, verify: 160 files, no problems): the sealed r34 pilot (117 members), the rewrite's final m4 test build (index-Csq4Zpws.js) and runner/perf-farm.mjs; the look runner and dev build are gone (D-18 settled). Step 3 (X-04) records stableMs for both scenes beside readyMs and declaredReadyMs; step 5 is X-05 as SPEC asks (tiers high, balanced and economy, both backends, living-orbit, living-route, walk and tractor-drive, --seconds 60 --runs 5, 120 files, about three hours), and compare adds x05 with the median and range per case, SPEC 20.2's 60 FPS target and the proposed hitch rule's verdict. Local plumbing check on the m4 kit (headless, numbers discarded, CPU 13.6 % before the run): the smoke passes all 13 checks, including the first stable frame recorded for both scenes; verify afterwards finds the 160 files unchanged, and the empty tmp/ the smoke leaves was removed (smoke, log, verify).
X-11 and X-12 pass (scripts/run-farm-x11-x12.ts --label m4 --alloc, headless 1920 × 1080 on owned ports, bytes and memory only; bytes, allocation, allocation sites, log). X-11: the staged r34 pack is 34 files of 7,326,021 B, its 23 GLBs 7,098,444 B (SPEC's 7,070,552 B is r33's); the public m4 chunk is 1,581,284 B / 462,421 B gzip, 91.30 % / 91.36 % of the Farm's D-15 ceiling. Before onReady the public page makes 29 requests with 8,906,023 B of bodies (page 1,858; code 1,581,284; JSON 56,437; GLBs 7,098,444; the grass binary 168,000; 8,914,723 B with headers); the test build makes the same 29 at every tier on both backends (8,918,214 B; its chunk is 12,191 B larger); nothing is fetched in the 5 s after ready in any load. X-12: after a forced collection before and after, the used heap grows by 0.074 to 0.172 MB over 1,000 frames in all eight cases (WebGPU and forced WebGL2, tiers high and minimal, living-orbit and walking; camera moving, Rowan on foot, tier held), against a 2 MB limit. The sampled allocation sites beside it are a diagnostic, not a rule.
Tablet (M4): minimal on WebGPU, X-07 and the governor below 20 FPS
Session (2026-09-30, 04:20 to 04:30 UTC; the device kit rebuilt from the m4 outputs, run-tablet.ps1 -Steps state,tier,soak; log, summary). Same device, Chrome 154.0.8037.57, panel 90 Hz, and the M3 method and rules (below): adb client only, the forward and reverse lists empty before and after every run, nothing installed, no setting changed. Device state: thermal status none (0) throughout; battery 93 to 92 % while charging on USB; AP 23.5 °C at the first state check (the screen was dozing), 34.2 °C after the tier runs and 34.5 to 39.1 °C during the soaks; brightness on automatic, 56 at the first state check and the WebGPU tier run and 51 from the WebGL2 tier run on. This PC's load before each run (13 to 44 % CPU) is recorded beside it; it affects only the time to ready (indicative: 7.3 to 9.3 s), not the tablet's frame intervals.
Tier (FARM-007 on the device; WebGPU, WebGL2): the same adapter strings as in M3 (WebGPU vendor arm, architecture valhall, device and description empty; WebGL2 ANGLE (ARM, Mali-G68, OpenGL ES 3.2)) now give minimal on both backends, with the reasons tablet / mid GPU; Mali-G68 class on WebGPU (X-07) and tablet / low GPU; Mobile WebGL2 startup mitigation (D-03), equal to classifyDevice on the page's own probe.
Soaks (soak); frame intervals are the tablet's own:
| Run | Tier | Length | Frames | Median | p95 | Mean FPS | Frames over 50 / 100 ms (longest) | Tasks over 100 ms | Governor |
|---|---|---|---|---|---|---|---|---|---|
| living-orbit | minimal (assigned) | 210 s | 12,431 | 16.6 ms | 17.3 ms | 59.2 | 2 / 0 (91.6 ms) | none | 0 changes |
| living-orbit | high (forced) | 210 s | 3,791 | 50.2 ms | 68.3 ms | 18.0 | 1,981 / 30 (181.3 ms) | none | 3 changes, 1 reversal, 1 lock; final level 1 (grass .75) |
| living-route | minimal | 60 s | 3,613 | 16.6 ms | 17.5 ms | 60.2 | 0 / 0 (33.9 ms) | none | 0 changes |
| tractor-drive | minimal | 60 s | 3,612 | 16.6 ms | 17.2 ms | 60.2 | 0 / 0 (31.3 ms) | none | 0 changes |
- X-07 passes SPEC 20.2's tablet row at the tier the owner assigned (30 FPS or better with p95 under 33.3 ms in orbit and in play, no task over 100 ms after ready): about 60 FPS at one vsync in every run. The orbit's two frames over 50 ms fall at 103.1 s (91.6 ms) and 125.2 s (51.2 ms); none in the first two revolutions. The M3 runs at economy (before the warm pass) are retained below; the tier differs, so the two sets are not a before/after of the M4 fix. The comparison still missing is a same-session pilot run on the tablet; SPEC 20.2's table still names economy as this class's expected tier.
- FARM-004 on the device: forced to high, the governor now acts below 20 FPS: one step down at 8.3 s (grass .75), reverted by the benefit rule at 13.8 s (the step bought less than 8 % of the median), a second step at 74.3 s that locks (the second reversal); 0.86 changes a minute, one lock, so X-10 holds. M3's same run made no change at 17 FPS. The lock stops it above the ladder's floor, so it never advises a lower tier, and an explicitly forced high stays at 18 FPS; that is SPEC 8.3's reversal rule, unchanged (an owner question below). Its 30 frames over 100 ms fall in the first 0.7 s after ready (5), at 28.4 s and 103.0 s, and in 128.3 to 138.2 s (23), with no long task: the GPU, not the main thread, is the limit there.
- Hitch sample for a mobile row: at the assigned tier the worst 60 s window had 2 frames over 50 ms and none over 100 ms. The hitch README gives this as a candidate row for the coordinator; one device and one session do not make it a proposal.
Tablet (M3): the owner's Galaxy Tab S9 FE
Device and method. Samsung Galaxy Tab S9 FE 5G (SM-X518U, serial R52X405L12T, Android 16, Exynos 1380 with a Mali-G68 GPU), Chrome 154.0.8037.57, panel at 90 Hz (display.renderFrameRate 90), screen 823 × 1317 CSS px at DPR 1.75 (page viewport 823 × 1142). It is driven from this PC over USB with the device kit (packages/farm/dist/device-kit/: the final m3 test and dev outputs copied unchanged, check-device.ps1, run-tablet.ps1, README, MANIFEST.json with SHA-256 per file). Each run serves the scene from its own server on 127.0.0.1:4400–4499, maps that port with adb reverse, reaches the tablet's Chrome DevTools with adb forward on 4495–4499, removes exactly the mappings it made (the lists before and after are recorded; the Golden Gate builder's 4600–4649 were never touched) and uses the adb client only. Nothing was installed on the tablet beyond the pages its Chrome opened, and no device setting changed. The device state is recorded before and after every result (battery level, temperature and charging state; HAL temperatures; thermal status; brightness and mode; wakefulness; display mode): the thermal status stayed none (0) throughout; the battery went from 79 to 73 % while charging on USB power; the AP temperature ranged from 22.1 °C (the first state check) to 50.2 °C (the end of the full-frame look captures); the brightness stayed on automatic, 45 at the first state check and at the start of the first tier run (23:16:14 to 23:16:19 UTC) and 56 in every record from 23:16:32 UTC on. Results: evidence/tablet-2026-09-29/.
Assigned tier and adapter strings (WebGPU, WebGL2):
| Backend | Adapter string read on the tablet | Classification | Tier |
|---|---|---|---|
| WebGPU (default) | adapter info vendor arm, architecture valhall (device and description empty) |
tablet, mid GPU | economy (render scale .75, grass .25) |
WebGL2 (?backend=webgl2) |
unmasked renderer ANGLE (ARM, Mali-G68, OpenGL ES 3.2), unmasked vendor Google Inc. (ARM) |
tablet, low GPU, plus the mobile WebGL2 startup mitigation (D-03) | minimal (render scale .6, grass 0) |
Both equal classifyDevice on the page's own probe with the Farm tier table. GG-008 (Samsung Xclipse) is neither triggered nor resolved: this tablet reports Mali strings. Time to ready on the tablet (indicative only; the pack is served from this PC, whose load sample is recorded beside it): 8.5 s on WebGPU and 7.0 s on WebGL2.
B-09 with real touch passes 14 of 14 steps on both backends on the final build (see Validation); the safe-area and mouse-layout steps are emulation-only and pass in the emulated set.
Governor soak and the SPEC 20.2 tablet proposal (soak). Frame intervals are the tablet's own, valid with the device state beside them:
| Run | Tier | Length | Frames | Median | p95 | Mean FPS | Frames over 50 / 100 ms | Tasks over 100 ms | Governor |
|---|---|---|---|---|---|---|---|---|---|
| living-orbit | economy (assigned) | 210 s | 6,335 | 33.2 ms | 49.8 ms | 30.2 | 215 / 6 | 2 (102, 101 ms) | 0 changes, 0 reversals, 0 locks |
| living-orbit | high (forced) | 210 s | 3,570 | 65.2 ms | 68.1 ms | 17.0 | 2,229 / 12 | none | 0 changes: below 20 FPS it never evaluates (FARM-004) |
| living-route | economy | 60 s | 2,004 | 33.0 ms | 34.3 ms | 33.4 | 17 / 3 | none | 0 changes |
| tractor-drive | economy | 60 s | 1,654 | 33.4 ms | 51.3 ms | 27.6 | 288 / 2 | 1 (156 ms) | 0 changes |
X-10 (at most four changes a minute and one lock) holds at the assigned tier. SPEC 20.2 proposes 30 FPS or better with p95 under 33.3 ms in orbit and in play and no task over 100 ms after ready. Orbit meets 30 FPS but not the p95 or the task rule; the route meets 30 FPS and the task rule, with p95 at 34.3 ms; tractor-drive misses all three. X-07 is recorded as failing that proposal. The pilot's recorded tablet numbers (34.2 FPS orbit, 44.7 FPS route, 83 ms worst gap) come from an earlier session, so a same-session pilot run is the comparison still missing. Chrome's frame intervals on this tablet are whole vsyncs (steps of 16.6 ms, although the panel reports 90 Hz): a median of 33.2 ms is two vsyncs, and a p95 under 33.3 ms needs nearly every frame within two vsyncs. FARM-004 (KIT-REQUESTS.md) records the governor gap for the owner: SPEC 8.3 evaluates only with 60 samples in its 3 s window, so a device that starts too high and runs below 20 FPS never steps down; the kit is unchanged.
Where the long frames fall (from the kept interval series; the frames-over columns were added after the owner's report of WebGL2 hitches at 20:35, from the same files). In the economy orbit, 5 of the 6 frames over 100 ms fall 5.3 to 6.2 s into the soak and the sixth at 102.5 s, and frames over 50 ms per 35 s revolution run 116, 35, 22, 32, 10 and 0; forced high has 5 in 5.3 to 6.6 s and 7 in 72.5 to 101.8 s; the route's 3 fall at 26.8 to 27.5 s and the tractor's 2 in its first second. The first revolution carries most of them. A no-timing count on this PC (m3 test build on forced WebGL2 at minimal, two living-orbit revolutions after ready; counts) finds 8 WebGL program links in the first revolution (4 as the orbit starts, 4 near 80°; pipelines 166 to 174) and none in the second, no texture, geometry or render-target allocation and no governor change in either, and three r186's WebGL backend re-specifying each uniform buffer with bufferData on every update in both. Lazy shader compiles as objects first enter view can therefore explain first-revolution hitches only; later ones need another cause (candidates: that per-draw bufferData churn, garbage collection, and on this PC its load). The owner's report is the first M4 item; M3 records observations and makes no fix.
WebGL2 play positions (yard, house porch, bridge) were captured on the tablet as screenshots: its own view at the minimal tier with the HUD, and the 1280 × 720 parity framing at tier high (device view, parity framing). They are not compared with the pilot.
Look options on the tablet: see the Look exploration section. Every tablet result's device state and mappings are in its own file; the state checks are under evidence/tablet-2026-09-29/state/, and each step's summary is summary-<steps>[-<suffix>].json.
three 0.186.0 evidence
| Check | Result |
|---|---|
| B-00a | PASS: async initialization and visible cube |
| B-00b | PASS: WebGPU backend and adapter reporting |
| B-00c | PASS: forced WebGL2 rendering |
| B-00d | PASS both backends: explicit shadows; omitted-prop reset documented and supported PCF type normalized |
| B-00e | FAIL both backends: 8/144 tiles exceed .02; worst tile mean .1184. Required Z4 startup-only DPR fallback applied |
| B-00f | PASS: outer StrictMode replay, fresh remount resources, late-init disposal. Nested R3F replay was not observed; the initial overstrict fixture failures are retained |
| B-00g | PASS: terminal counters, owned device/context release, no render after disposal |
| B-00h | PASS: Neutral/WebGPU, Neutral/WebGL2, AgX/WebGPU, AgX/WebGL2; distinct HDR patches and close cross-backend pixels. ACES reference also renders |
M0 raw results. Neutral remains the default for later scenes; the Farm keeps ACES Filmic at exposure .95 (D-18).
| Rule | Evidence and remaining limit |
|---|---|
| Z1 | Exact three 0.186.0, revision-186 guard, pin check and one-copy source consumer |
| Z2 | Derivative assertion and demo normal-map proof pass; Farm woodland tangent-safe A/B passes on both backends (D-06); batch derivatives equal the pilot's 11 plus one woodland |
| Z3 | Terminal ownership, late initialization and GPU release proven; B-05 passes for the kit demo and for the Farm on both releases and profiles under C-03/C-04 |
| Z4 | Startup-only pixel ratio; every live governor DPR step removed; the Farm live ladder steps grass only (R5-01); B-13 passes on the full scene on both backends (M3b: a governor grass step plus a resize equals a fresh load, 144/144 tiles at eye-height and the watermill, stream crop included) |
| Z5 | No velocity/MRT/Loop qualification claimed for public use. The M3 dev-build look exploration uses an emissive MRT (bloom), a velocity MRT (TRAA) and a normal/depth pre-pass (GTAO) inside one RenderPipeline; all 12 presets build and render without console errors on both backends and their captures are on the contact sheet, but nothing leaves the dev build and nothing becomes a default (D-18) |
| Z6 | Explicit aliases to the same WebGPU build; one installed/bundled three package |
| Z7 | Field Grass vendored with source, licence and provenance; clock and disposal tests pass; 14,000 tufts in 16 meshes at every static pair, and B-06 passes on both backends, including the stream crop |
M4 findings on three 0.186.0 (no dependency or pin change; hitch notes): an instanced mesh with at most 1,024 instances keeps its matrices in a uniform buffer declared in its own vertex shader, so each compiles its own program or pipeline, lazily when it first comes into view (on WebGL2 a synchronous link of about 100 ms inside that frame), and uploads the whole buffer every frame it draws; a DynamicDrawUsage attribute is uploaded every frame whether or not it changed; and compileAsync builds node materials in slices across tasks, so many concurrent builds carry their temporaries into the old generation (about 250 MB with all 221 chains at once). The Farm avoids all three through documented API (R6-01, R6-03, R6-04); FARM-008 tells the other scenes.
Decisions and deviations
The full dated log is in PROGRESS.md; DECISIONS.md governs where it differs from SPEC.md. M1 work: C-01/C-02 (governor floor and trace), C-03/C-04 (listener exception, heap protocol), R2-02 to R2-14 (renderer hosting, Clock facade, owned Windows device, entry conditions, scoped Fiber errors, r186 texture cleanup, streamer bound), R3-02 (D-15 baseline recovery). No dependency file or pin changed.
Deviations since M2a, each bounded and recorded without waiving an assertion:
- R3-08: the SPEC counts mix 96 m and historical 24 m values; the current sealed values are gated, the historical ones recorded.
- R4-01: overview minimum distance 3 (INV S23), with a closer named view relaxing it; supersedes M2a-FIX-01's constant 0.
- R4-02: follow camera through the kit
FollowRig,hideSubjectBelow = 1.25 - .14. - R4-03: play flag and Escape veto; a page-initiated stop dismounts at the first exit option (no pilot equivalent).
- R4-04: Rowan animates in play without ambient animation; doors snap under reduced motion.
- R4-05: the kit
fixedSteptolerance of 1e-12 instead of an exact comparison (never diverged in 4,291 oracle frames). - R4-06: the ground collider is the prepared terrain soup on an off-scene proxy, registered in the pilot's order.
- R4-07: parity phase calibration for stage m2d (the rewrite takes the pilot's shader-clock phase bracketed around its capture) and outline normalization in the capture CSS.
- R4-08: hub run kit with ports of the pilot's workloads in test and dev builds only, the kit's
cpuRenderTimes(), SPEC 20.3's runner flags, the 30 s and no-walk deviations above, the engagement check and the pinned headed window. - R4-09: the r34 run compared its colliders with the only frozen (r33) fixture; the sealed r34 collision world is now frozen from the sealed r34
play.mjs(soups byte-identical; +264 static triangles, all in the r34 farmhouse), the parity script selects the fixture by release, and the r34 attempts were re-evaluated offline from their recorded counts, with no new capture. - R4-10: under D-20, P-16 deferred (B-10 is M3b), P-19 dropped under the cut line (D-09), B-13 to M3b, WebGL2 play positions not captured, P-31 to P-35 left to M3; and the owner's rules on PC timing and headed windows (11:35, 11:40).
M3 (round 5), each bounded and recorded in PROGRESS.md without waiving an assertion:
- R5-01: the Farm live ladder steps only the grass fraction (Z4 removes pixel-ratio steps and the Farm has no other SPEC 15 knob); a one-level tier never steps, advises or persists a lower tier.
- R5-02: D-22 on the Farm: on touch the overview camera pad goes (drag orbits, pinch zooms, two fingers pan) and one Reset view target stays; in play only the joystick and at most one context button; keyboard and mouse unchanged.
- R5-03: HUD placement in Farm-scoped CSS (a static toolbar in a full-HUD layer, the GG-007 workaround; camera controls bottom-left; the status line never takes input and moves clear of the joystick, the context button and Rowan). This fixes the site's 390 px overlap.
- R5-04: B-09 pinch limits are checked on the stored follow offset (1.6 to 10 m); the camera sits .14 m inside it in the open, as in the pilot.
- R5-05: the B-09 runner spaces a tap at least 600 ms after a drag, pinch or stick hold, because Chrome suppresses a tap that lands just after a gesture ends (input spacing, not a measurement).
- R5-06:
bun run lintisscripts/lint.tson the pinned TypeScript alone (no linter pin exists); gated on this builder's packages, advisory elsewhere. - R5-07: tablet runs through the device kit with real touch (CDP
Input.dispatchTouchEventinto the device's Chrome, no emulation); the safe-area and mouse-layout steps are emulation-only and skipped there; the runner clears its own origin's storage on the device before loading. - R5-08: tablet frame time is the tablet's own and comes in whole vsyncs; each look cell is one sample, so a one-vsync difference between neighbouring cells is within its spread; the tablet's time to ready is indicative only.
- R5-09: look captures use the B-06 framing (full-frame 1280 × 720 canvas, page shell and HUD hidden); the first, shell-framed captures are retained; frame time runs on the standalone page with its shell.
- R5-10: the B-08 runner selects the collider fixture by the build's pack release (R4-09).
- R5-11: U-18 regression fixed: the pilot's touch tractor status restored verbatim (SPEC 13.1, INV 6); the outputs rebuilt; the rebuilt chunks are byte-identical to the earlier ones apart from that literal and the chunk names, so the timing and capture runs on the earlier build stand; the counted functional runs were repeated.
- R5-12: X-04 attempt: a measurement-only change (the hub runner records
declaredReadyMs, compare writesx04Declared); no pipeline pre-compilation (SPEC 6.4) and no change to the onReady moment (SPEC 5.2(3)); both are owner decisions. - R5-13: FARM-004: the governor cannot evaluate below 20 FPS under SPEC 8.3; recorded for the owner, kit unchanged.
- R5-14: FARM-005:
scripts/static-server.mjsserves only when it, or a build output'sserve.mjscopy, is itself the process entry (isServerEntry), so a bundle that inlines it (the hub kit'srunner/look-farm.mjs, through scene-kit'sserveOwned) no longer starts a stray server on 4400 or stays alive; regressionscripts/tests/static-server.test.ts(it failed before the fix); the m3 outputs'serve.mjskeep the earlier guard, which is correct for direct use; the hub kit is rebuilt. - R5-15: the owner's note of 20:35, relayed by the coordinator: every frame measurement keeps its raw interval series and counts frames over 100 ms beside p50, p95, p99, max, frames over 50 ms and long tasks (hub runner, look runner, tablet soak); no attribution counters; the WebGL2 hitches are the first M4 item, and M3 records a no-timing observation only.
M4 (round 6), each bounded and recorded in PROGRESS.md without waiving an assertion:
- R6-01: the SPEC 6.4 warm pass (owner decision 2026-09-29 22:05, D-23): the first world of a mount stays hidden behind the loading screen while three's
compileAsynccompiles every drawable below the world root (culling off, 8 chains at a time,WARM_CONCURRENCY, chosen by measurement), then one frame draws it unculled (the shadow and first-draw variants three skips while pre-compiling), then culling is restored and the scene declares ready. A world rebuilt after a tier change warms while it stays visible. SPEC 6.4 and 5.2(3) carry the owner's wording and the dated note. - R6-02: X-04 is the time to the first stable frame for both scenes (the first frame from the ready frame on with no shader, program or pipeline creation, no asynchronous creation in flight and at most 50 ms); the hub runner, its
compareandscripts/run-farm-x04.tsmeasure it the same way and keep the ready and declared moments as extra columns (x04Ready,x04Declared). - R6-03: the Farm reads instance matrices as instanced attributes through three's documented
renderer.debug.onNodeBuilderCreatedhook instead of three's per-mesh uniform buffers (packages/farm/src/world/instance-attributes.ts); test and dev builds keep the uniform path behind?instanceUniforms=1for A/B. The look A/B passes (see the M4 section). - R6-04: the Farm's dynamic batches take static usage (
optimize.ts); with FARM-006 only a batch whose source moved is uploaded. - R6-05: FARM-004, as TASK-M4 item 2 approved: the governor evaluates once its window holds 60 samples or has spanned
windowMs, whichever comes first; the cadence, thresholds, benefit check, reversal lock and advice are unchanged. - R6-06: X-07 (owner decision 22:10, D-25, FARM-007): the Tab S9 FE's GPU class on WebGPU starts on minimal, its WebGL2 tier; the rule is narrow (a named Mali-G5x or G6x, or vendor
armwith architecturevalhalland no model name) and its adapter strings are recorded beside it. - R6-07: FARM-006 (kit, additive behaviour, no API change): a dynamic batch's matrices are written, marked for upload and re-bounded only when one of its source matrices changed.
- R6-08: the proposed hitch rule for X-05 is reported by the hub kit's
compare(x05[].proposedHitchRule) and adopted by no gate; the coordinator decides. - R6-09: the hitch method: headless 1920 × 1080 at DPR 1 on an owned port, three living-orbit revolutions from the first frame after ready, a plain and a traced run per case; a trace's collection statistics cover the whole trace, including the time after the last frame, so each major collection was located against the frame series before it was attributed.
- R6-10: FARM-008 is an advisory for the other scenes (no kit change); the Farm keeps its two measures in
packages/farm.
No new third-party exception since M2a: C-03 remains the only infrastructure exception, and the vendored Field Grass patches (R3-04) are unchanged. Kit changes in M3 are compatible additions recorded in KIT-REQUESTS.md (FARM-002 launchChrome window size, FARM-003 the test-only adapter-string hook, GG-010 chase pinch zoom and pinch-drag guard, FARM-005 the static server's entry guard). M4 adds FARM-004 (the governor's elapsed-window rule), FARM-006 (dynamic batches written only when a source moved) and FARM-007 (the X-07 classification), each done with its test first, and files FARM-008 as an advisory; no kit API changed.
Bundle sizes, hygiene and licenses
| Output | JavaScript bytes | gzip bytes | Evidence |
|---|---|---|---|
| Kit demo public | 1,418,364 | 412,484 | public |
| Kit demo test / dev | 1,427,698 / 1,429,642 | 415,740 / 416,340 | test, dev |
Farm public r33 and r34, index-C30lhJPq.js (identical chunk) |
1,574,581 | 460,130 | r33, r34 |
Farm test r33, index-CQHeHCPN.js (before the hub-kit workloads) |
1,579,935 | 462,087 | r33 test |
Farm dev r33, index-C9CkQbwt.js (before the workloads) |
1,581,927 | 462,778 | r33 dev |
Farm test r34, index-D4tLUAUg.js (with the workloads; the hub kit's build) |
1,586,258 | 464,174 | r34 test |
Farm dev r34, index-CkVTJfoR.js |
1,588,250 | 464,845 | r34 dev |
Farm public m3 (r34 pack), index-CavZFEhs.js |
1,577,844 | 461,199 | m3 public |
Farm test m3, index-y-h2nCPw.js (the hub kit's and device kit's build) |
1,589,805 | 465,377 | m3 test |
Farm dev m3, index-B1rTJVXs.js + FarmLook-TdozexLW.js (look exploration, loaded only behind KILN_DEV) |
1,596,319 + 84,983 = 1,681,302 | 467,634 + 46,964 = 514,598 | m3 dev |
Farm public m3 in the site's resolution (without the kit's three facade alias; scratch build, deleted after measuring) |
1,931,509 | 545,492 | double import |
Farm public m4 (r34 pack), index-Cblb5RZg.js |
1,581,284 | 462,421 | m4 public |
Farm test m4, index-Csq4Zpws.js (the hub kit's, the device kit's and the after2 hitch runs' build) |
1,593,475 | 466,666 | m4 test |
Farm dev m4, index-DOAXpML5.js + FarmLook-W7_nu9PA.js (look exploration, only behind KILN_DEV) |
1,599,989 + 84,983 = 1,684,972 | 468,935 + 46,964 = 515,899 | m4 dev |
Farm public m4 in the site's resolution (without the kit's three facade alias; scratch build, deleted after measuring) |
1,934,965 | 546,755 | double import |
D-15 as amended by the owner at 12:45 (2026-09-29): the ceiling is per scene. The kit's stays at 1,557,399 B / 452,687 B gzip, ceil(1.1 × the recovered kit demo baseline of 1,415,817 B / 411,533 B, R3-02), and the kit demo passes. The Farm's is its M2 public chunk plus 10 %: base 1,574,581 B / 460,130 B, ceiling 1,732,040 B / 506,143 B gzip. The final m3 public chunk is 1,577,844 B / 461,199 B gzip, 91.10 % / 91.12 % of the Farm ceiling (+3,263 B / +1,069 B gzip over the M2 chunk, for the D-22 HUD, the tier table and ladder and the M3 UI). The look exploration is absent from public and test outputs (B-12, hygiene). Classic WebGLRenderer double import (TASK-M3 known defect, D-15): the Farm's own build has none: every bare three import resolves to the kit's facade over three/webgpu, so three.module.js is not in its module graph. The site's resolution keeps only the exact @kiln-scenes aliases and dedupe, so three.module.js joins three/webgpu and adds 353,665 B / 84,293 B gzip, putting the chunk at 111.52 % / 107.77 % of the Farm ceiling. The Farm build cannot avoid that from its side; the remedy is one three alias in the site (the kit's sceneSourceConfig already carries it), for the coordinator (evidence). The workloads exist only in test and dev builds; the public chunk is identical with or without them. Earlier Farm test measurements remain in evidence/m2/bundles, the m3 builds' in evidence/build/m3/bundles, including the pre-revert chunks and the revert diff.
Staged Farm packs: 23 GLBs of 7,070,552 B (r33) and 7,098,444 B (r34), 7,310,663 B and 7,338,680 B with data, licences, pack.json and SHA256SUMS; only farmhouse.glb changes. Verified pack comparison. The kit demo's hygiene and licences pass; the Farm's B-12 public/dev hygiene passes on both releases and on the m3 outputs, and the public m3 output passes the bundle hygiene check (log). M4: the public m4 chunk is 1,581,284 B / 462,421 B gzip, 91.30 % / 91.36 % of the Farm ceiling (+3,440 B / +1,222 B over m3, for the warm pass, the instanced-attribute hook, the static batches and the three kit changes); in the site's resolution it would be 1,934,965 B / 546,755 B, 111.72 % / 108.02 % (+353,681 B / +84,334 B gzip), so the site's three alias is still needed (evidence). The public m4 output passes the bundle hygiene check (3 files, 1,588,779 B, no forbidden token; log) and B-12 on both profiles. The production licence audit with the public m4 module graph (146 bundled modules) passes: 19 packages (12 production, 6 bundled, 1 vendored), all MIT or BSD-3-Clause; @react-three/fiber ships no separate licence file and its declared MIT licence is recorded in the notices (log). Every m4 output carries THIRD-PARTY-NOTICES.txt (22,231 B) rendered from that audit by the build. check:site-compat was not re-run: it re-checks the kit demo's consumer and would overwrite the M1 evidence of B-14, whose Farm consumer is the site round's (TASK-M4: not M4). Project code is MIT, Copyright (c) 2026 Matthew Kissinger; the staged packs keep the runtime asset licence files and sealed notices byte for byte.
Look exploration (M3 item 6)
D-18 stands, and nothing here is a default. The exploration lives only in the dev build: packages/farm/src/look/ (options.ts, pipeline.ts, FarmLook.tsx) is loaded by a dynamic import behind KILN_DEV, so the public and test outputs never contain it (B-12 hygiene; the dev output's FarmLook chunk is 84,983 B / 46,964 B gzip). Options come from ?look= (a preset name or a list such as fxaa,ao,bloom or neutral,exposure=1.2), from the developer panel or from the setLook test hook; lookState() reports the requested and effective options, the passes, the MSAA samples, the exposure and anything a tier disabled. The default renders exactly as before: straight to the canvas with the renderer's 4× MSAA and ACES Filmic at .95, with no pipeline. Every other option builds one three 0.186 RenderPipeline (pipeline.ts) in this order: scene pass (with MRT outputs where needed), GTAO, bloom, TRAA or SMAA on linear colour, tone mapping and sRGB once, grade, vignette, FXAA on display colour. AO is off on the economy and minimal tiers (the tablet's), vignette and grading are off by default, and all 12 presets build and render on both backends without browser errors (Z5).
| Preset | Passes |
|---|---|
default |
Direct render with the renderer's 4× MSAA, ACES Filmic .95 (today's look; no pipeline) |
aa-none |
Scene pass without MSAA |
fxaa |
Scene pass without MSAA; FXAA on display colour |
smaa |
Scene pass without MSAA; SMAA (1×, medium) on linear colour |
traa |
Scene pass without MSAA, with a velocity output; TRAA |
ssaa |
SSAA scene pass (four jittered renders) |
ao |
4× MSAA scene pass; a normal and depth pre-pass and GTAO at half resolution on indirect light (off on economy and minimal) |
bloom |
4× MSAA scene pass with an emissive output; bloom on the emissive output only, never on lit colour |
vignette |
4× MSAA; vignette after tone mapping |
grading |
4× MSAA; a mild grade after tone mapping (per-channel slope and power, saturation 1.08) |
all |
SMAA, GTAO, bloom, vignette and grade |
neutral |
4× MSAA; Neutral tone mapping at 1.3316, the exposure that maps 18 % grey as ACES does at .95 |
Contact sheet: evidence/look/contact-sheet.png, with every number in contact-sheet.json (scripts/make-look-contact-sheet.ts --look-ab). Its rows are the PC at high (hero, fences, house porch, watermill wheel) and the tablet at its assigned economy tier (hero, house porch), then each option's frame time per device and tier (PC, tablet WebGPU, tablet WebGL2) and a pending laptop row. Captures use the B-06 framing: a 1280 × 720 full-frame canvas at DPR 1 with the page shell and HUD hidden, the clock frozen at time 0 and ambient animation off (R5-09). The PC captured all 12 B-06 views at high and four at economy (PC); the tablet captured four views at economy through a 1280 × 720 device-metrics override (tablet). The first captures, which included the page shell, are retained (PC, tablet). D-18 pairs: the three Farm views (hero, house porch, props) in ACES .95 and Neutral 1.3316 on WebGPU and forced WebGL2 (WebGPU, WebGL2, index, view choice). The owner chooses; the default does not change.
What each option changes (data): each capture against the default capture of the same view, tier and device, as the mean absolute difference of the 8-bit channels and the share of pixels whose luminance moves by more than 8. This is a size of change for reading the sheet, not a gate.
| Preset | PC high, 4 views: mean / pixels | Tablet economy, 2 views | Where it shows |
|---|---|---|---|
aa-none, fxaa, smaa, traa |
1.0–7.4 / 1.5–23.3 % | 0.8–3.0 / 1.1–7.8 % | Edges; most on the fences' thin rails |
ssaa |
0.5–2.3 / 0.2–4.6 % | 0.4–1.0 / 0.1–0.9 % | Closest to MSAA |
ao |
2.3–7.4 / 6.6–27.0 % | 0 (off on economy) | Contact shadows: the 5th percentile of display luminance falls from .32 to .23 at the fences and from .38 to .26 at the watermill; 23 % of the porch's pixels change |
bloom |
0–3.6 / 0–11.3 % | 0–0.7 / 0–0.1 % | Only the stream water, whose material has an emissive term (11 % of the watermill view); the porch is unchanged |
vignette |
9.4–13.4 / 37.0–47.4 % | 9.4–13.4 / 37.1–47.5 % | The frame's edges, by design |
grading |
6.8–7.9 / 2.4–11.7 % | 6.8–8.0 / 1.8–8.3 % | Everywhere, slightly: warmer (mean red minus blue +.07) and more saturated (mean saturation +.06), on a 0–1 display scale |
all |
15.7–20.0 / 45.5–63.0 % | 11.3–16.4 / 31.3–53.8 % | The sum |
neutral |
7.8–12.8 / 26.2–60.7 % | 9.9–12.8 / 54.3–60.0 % | The tone curve: brighter highlights (the 95th percentile of display luminance from .81–.86 to .90) and more saturation (+.07 to +.10) |
Frame time per option (p50 / p95 ms; for the tablet also the mean FPS). Living-orbit with ambient animation on and the governor held by an empty synthetic trace; per preset, a 2.5 s warm-up after its pipeline is built, then a 6 s (PC) or 8 s (tablet) sample; one page per tier, presets in the order shown. The PC columns are indicative only: shared PC, headless, frame pacing uncapped (--disable-gpu-vsync --disable-frame-rate-limit), load sample per tier block (CPU 1.5 to 24.2 %, GPU 3D 1 to 22 %). The tablet's numbers are its own, vsync-bound, with its device state before and after every tier block: thermal status none, battery 79 to 77 % (WebGPU) and 76 to 73 % (WebGL2) while charging, AP 27.4 to 45.0 °C and 27.9 to 50.1 °C, brightness 56 on automatic. The table shows each backend's assigned tier and high; all four tiers are on the sheet and in its JSON (PC runs, tablet WebGPU, tablet WebGL2).
| Preset | PC high* | PC economy* | Tablet WebGPU economy (assigned) | Tablet WebGPU high | Tablet WebGL2 minimal (assigned) | Tablet WebGL2 high |
|---|---|---|---|---|---|---|
default |
7.1 / 7.9 | 6.6 / 7.4 | 33.3 / 50.2 (25.5) | 65.7 / 83.3 (16.0) | 16.6 / 16.7 (52.6) | 66.3 / 100.0 (16.3) |
aa-none |
7.3 / 8.2 | 6.8 / 7.6 | 33.2 / 33.3 (34.2) | 49.8 / 50.9 (22.3) | 16.6 / 16.7 (60.2) | 33.2 / 66.4 (29.4) |
fxaa |
7.4 / 8.8 | 7.0 / 8.1 | 33.2 / 33.4 (30.3) | 49.8 / 51.2 (22.1) | 16.6 / 16.7 (54.4) | 33.2 / 66.5 (27.7) |
smaa |
7.6 / 8.6 | 7.0 / 7.9 | 33.2 / 49.7 (30.0) | 49.8 / 51.8 (20.0) | 16.6 / 16.7 (57.2) | 33.9 / 65.8 (25.0) |
traa |
9.9 / 10.5 | 9.4 / 10.4 | 49.9 / 69.0 (17.6) | 82.8 / 83.2 (13.3) | 16.7 / 33.3 (39.3) | 49.9 / 133.3 (15.5) |
ssaa |
16.4 / 33.0 | 14.9 / 17.5 | 83.0 / 83.2 (12.7) | 166.1 / 166.2 (6.0) | 49.8 / 50.1 (22.1) | 99.7 / 133.4 (8.8) |
ao |
12.8 / 13.9 | 6.8 / 7.8 (AO off) | 33.2 / 33.3 (35.5) (AO off) | 83.0 / 83.1 (12.4) | 16.6 / 16.7 (59.9) (AO off) | 83.1 / 132.1 (11.9) |
bloom |
8.0 / 9.1 | 7.4 / 8.6 | 49.8 / 50.0 (22.6) | 83.0 / 100.0 (12.3) | 16.6 / 33.3 (49.3) | 83.0 / 116.7 (12.4) |
vignette |
7.5 / 8.3 | 6.9 / 7.8 | 33.2 / 49.9 (28.0) | 66.2 / 66.7 (16.9) | 16.6 / 16.7 (54.5) | 66.2 / 100.0 (16.7) |
grading |
7.5 / 8.4 | 7.1 / 8.5 | 33.2 / 49.9 (28.0) | 66.4 / 66.9 (16.6) | 16.6 / 16.7 (56.6) | 66.3 / 100.0 (16.6) |
all |
13.5 / 15.0 | 7.4 / 8.2 (AO off) | 33.9 / 50.0 (25.1) (AO off) | 99.6 / 149.7 (9.9) | 16.6 / 33.2 (53.0) (AO off) | 66.5 / 100.1 (13.0) |
neutral |
7.3 / 8.0 | 7.0 / 8.0 | 33.8 / 50.0 (25.0) | 52.4 / 66.6 (17.3) | 16.6 / 16.7 (59.8) | 66.1 / 100.0 (16.6) |
* Indicative only: this PC, under load, never a result.
How far a cell can be trusted. Each cell is one sample, and the tablet's intervals are whole vsyncs (16.6 ms), so a one-vsync step at p95 between neighbouring cells is within the spread (R5-08). The sheet measures the spread directly, because at economy and minimal the ao preset renders exactly the default (AO is off there and no pipeline is built). In each such pair on the tablet, the default cell, the first of its tier block and sampled in the orbit's first seconds on a freshly loaded page, holds the block's long frames (maximum 133 to 283 ms, 3 to 12 frames over 50 ms), and the ao cell about a minute later holds none (maximum 33.3 to 49.9 ms); the mean reads 25.5 against 35.5 FPS on WebGPU economy and 52.6 against 59.9 FPS on WebGL2 minimal; and where frames split between two vsyncs the median flips by one vsync for the same configuration (WebGL2 economy: 16.8 against 33.1 ms at 36.9 and 37.6 FPS). So at the assigned tiers the ao row, not the default, is the fairer reference, and only median differences of at least one vsync that recur across presets are findings. The long frames in each block's first cell follow the same pattern as the soak's first-revolution hitches (Tablet section).
Findings (the tablet's; the PC's only as indicative):
- At the tablet's assigned tiers the default is already vsync-bound in the median: two vsyncs (33 ms) on WebGPU economy and one (16.6 ms) on WebGL2 minimal.
aa-none,fxaa,smaa,vignette,gradingandneutralkeep that median;traaandbloomadd one vsync on WebGPU economy (49.8–49.9 ms) andssaaadds three (83 ms). On WebGL2 minimal onlyssaamoves the median (49.8 ms);traa,bloomandalllift p95 to 33.2–33.3 ms, andtraalowers the mean from 50–60 to 39 FPS. - At the tablet's high tier the renderer's 4× MSAA is the cost: the MSAA presets without further passes (
default,vignette,grading,neutral) sit at 52–66 ms (WebGPU) and 66 ms (WebGL2) in the median, whileaa-none,fxaaandsmaasit at 49.8 ms and at 33.2–33.9 ms. The tablet runs economy and minimal, not high, so this matters only if a higher tier were assigned to such a device; replacing MSAA with a post-process AA changes edge pixels (table above), so it would be a look decision for the owner. - GTAO and SSAA are the expensive options wherever they run (PC high, indicative: about +5.7 ms and +9 ms at the median; tablet high: 83 ms and 166 ms). AO stays off on the tablet tiers, as the task set it.
- Bloom changes only the stream water in these views and costs one vsync on the tablet's WebGPU economy, so it buys little there.
- Neutral tone mapping at the matched exposure shows no cost beyond the spread on any device; the choice is one of look (D-18).
Better implementations for performance that keep the look: none adopted in M3. Each candidate keeps the pixels but is either excluded by SPEC or needs a kit change that the coordinator schedules; no M3 gate needs one:
- Pipeline pre-compilation before onReady (
compileAsync; Dawn compilescreateRenderPipelineAsyncpipelines in parallel): the X-04 attempt names per-pipeline shader builds as the start-up cost, and the tablet's long frames fall in the first seconds after ready. SPEC 6.4 forbids prewarm (INV 4.3), so this is the owner's decision (open decision 1). - Shadow-map reuse: animated clips move Farm casters every frame while ambient animation runs, so updating the sun's map only on change pays only when ambient animation is off, paused or reduced; it needs a kit knob and a moved signal from the mixers, doors and tractor.
- Rendering on demand while the clock is frozen or paused: identical pixels; it needs a dirty flag in the kit loop.
- FARM-001 preparation in a worker (terrain and collision soups, woodland points, meadow pixels): shorter main-thread tasks before onReady; it needs the FARM-001 hook to accept a worker-backed task.
- On WebGL2, three r186 re-specifies each uniform buffer with
bufferDataon every update (Tablet section); updating in place would keep the pixels, but it is a three.js internal and the pins are fixed.
M4 adopted the first candidate after the owner's decision of 22:05 (D-23, R6-01: the warm pass before ready) and removed most of the per-frame uniform uploads the last one describes by reading instance matrices as instanced attributes (R6-03); see the M4 section. The others stay as listed.
Not proposed, because they change the look or the sealed pack: KTX2 or quantized textures, LOD, FSR upscaling, and a post-process AA in place of MSAA as a default.
Laptop: pending, for the coordinator. The hub kit carries the dev build and runner/look-farm.mjs (README step 6: every preset per tier on both backends during living-orbit, then the sheet's four views at high and economy, about 20 minutes per backend). Its results/look-<backend>/ directories come back unchanged under evidence/perf/hub-<date>/, and scripts/make-look-contact-sheet.ts --laptop evidence/perf/hub-<date>/look-webgpu fills the laptop row. This builder never connects to the hub.
Open decisions, ownership and command log
Settled since M3 by the owner's decisions of 22:05 and 22:10 and by M4: which moment X-04 measures and whether to prewarm (D-23: the first stable frame, with the warm pass before ready), D-18 (unchanged: MSAA 4x and ACES .95 stay the default on every tier, so the hub kit's look step is gone), X-07's target for this tablet (D-25: its class starts on minimal on WebGPU), the WebGL2 play positions (D-25: the mobile record under D-20), X-08 and X-09 (D-25: deferred, no phone), FARM-004 (done) and the WebGL2 hitches (cause found and fixed).
Open for the owner and the coordinator after M4:
- Hub runs of the m4 kit (coordinator): step 2 (X-01 on m4; the warm pass and the instanced attributes change the frame path), steps 3 and 4 (X-04 as the first stable frame; the M3 hub failure on WebGPU measured the old moment), step 5 (X-05, 120 files, about three hours) and
compare. X-06 needs a run without a forced tier, which the kit does not make. - The proposed hitch rule (coordinator): adopt, adjust or drop it (per 60 s run at every tier: 0 frames over 100 ms, 0 over 50 ms, longest at most 50 ms;
comparereports it asx05[].proposedHitchRule), and decide whether minimal joins step 5. A mobile row needs a phone and a second session; the tablet's first numbers suggest 0 over 100 ms and at most 2 over 50 ms per 60 s at the assigned tier. - SPEC wording (owner): SPEC 20.2's tablet row still names economy as this class's expected tier, and SPEC 20.1's X-04 row still says "time to onReady"; this builder edited only SPEC 6.4 and 5.2(3), as instructed.
- Advice from a locked governor (owner, SPEC 8.3): with FARM-004 a tablet forced to high steps down, is reverted by the benefit rule and locks at level 1 on the second reversal; advice fires only at the ladder's floor, so a tier too heavy for the device that was forced explicitly stays at 18 FPS. Explicit quality props win by design; decide whether a lock above the floor should also advise a lower tier.
- The m4 B-05 baseline (coordinator): about 0.83 MB above m3's on both profiles after the initialization cycles, retained once per page (the growth over the measured cycles is unchanged and passes); not traced further. A candidate is module-level state the warm pass fills.
- FARM-008 (coordinator): route the advisory to Golden Gate and Foundry Floor (three 0.186.0's per-mesh instance uniform buffers: a program per mesh compiled lazily in view, and a full upload every frame; the Farm's two measures).
- The site round (coordinator): the site's
threealias (without it the m4 chunk grows by 353,681 B / 84,334 B gzip to 111.72 % / 108.02 % of its ceiling), the island check, the D-11 mobile copy (B-09 passes), and B-14's Farm consumer. - X-07 comparison (owner): a same-session pilot run on the tablet; the pilot's tablet numbers come from an earlier session.
- Later actions, unchanged: D-12 uploads and CORS (owner), the optional Field Grass upstream URL.
Not run in M3 or M4: the Golden Gate contract suite. It serves on the Golden Gate builder's ports 4600 to 4649, outside this builder's 4400 to 4499, and that builder runs it.
Ownership (M3). Every owned browser and server was closed in finally and recorded in its run ledger. The M3 final audit (01:23 UTC on 2026-09-30, 21:23 EDT on 2026-09-29) found none of the 20 browsers recorded in this round's run ledgers still running, no workspace-scoped runtime process and no listener on ports 4400 to 4499 (audit; read-only OS metadata, no socket); the M2 and M1 audits are historical. The one stray listener of this round was this builder's own: the second server on 4400 that the hub kit's look runner opened before FARM-005 was fixed, stopped by the builder at 20:45 EDT. On the tablet the shared adb server was used as a client only (never started, killed or restarted); the runners woke the screen, started Chrome, opened and closed their own tabs through a DevTools forward and removed their own forward and reverse mappings, and at the close the device's forward and reverse lists are empty; nothing was installed on the tablet. No listener on ports 8000 to 8099 was contacted, probed or stopped, and nothing outside this workspace was started on this PC. No hub connection, commit, branch mutation, upload, publication or paid/live model call occurred. This builder did not edit DECISIONS.md (last written 16:27 EDT, before this round began at 17:09) or packages/golden-gate/.
Ownership (M4). Every owned browser and server was again closed in finally and recorded in its run ledger. The M4 final audit (04:45 UTC on 2026-09-30, 00:45 EDT) found none of the 8 browsers and 3 runner processes recorded in this round's ledgers still running, no process whose command line names this workspace and no listener on ports 4400 to 4499 (audit; read-only OS metadata, no socket). On the tablet the shared adb server, already listening, was again used as a client only; the runners worked as in M3 and removed their own mappings, and at the close the device's forward and reverse lists are empty; nothing was installed. No listener on ports 8000 to 8099 or in the other builders' ranges was contacted, probed or stopped, and the hub was not contacted. This builder made no commit, branch mutation, upload, publication or paid/live model call; HEAD moved during the round from 833f6da to 18dccdb through two Golden Gate commits by others that touch only Golden Gate paths and its hub evidence, and this round's changes are uncommitted in the working tree. Edits stayed in scope: packages/farm, the kit's quality/ (FARM-004, FARM-007) and, additively, instancing/core.ts (FARM-006; noted in KIT-REQUESTS.md, kit tests pass), the Farm's runners and checks under scripts/, the root PROGRESS.md, REPORT.md (Edit only) and KIT-REQUESTS.md, and SPEC.md 6.4 and 5.2(3) as the owner's decision of 22:05 instructed; not packages/golden-gate, packages/foundry-floor or the kit's testing/node.ts. DECISIONS.md was written at 22:37 EDT by the coordinator (D-23 to D-26), not by this builder. The cleanup deleted only the four paths TASK-M4 item 5 names (packages/farm/dist/m3-wip, .tmp/m3-pre-revert, .tmp/b09, .tmp/tiering: 298 files, 55,610,157 B, measured before deleting); this round's other scratch stays in the git-ignored .tmp/ (about 2.57 GB, mostly the runners' Chrome profiles and the raw hitch traces the hitch README cites) for the coordinator to keep or delete.
| Command group | Evidence |
|---|---|
| Pins, install and replay; toolchain checks | Pins, install, replay, toolchain |
spike:test and M1 qualification |
B-00, Profile 1, Profile 2, C-03 |
| Farm staging, world and placement units | Staging, world, placements |
| Farm builds | r33, r34, r34 with workloads |
| Units | Full CPU suite, after B-08 |
| Contract suite | r33 default, r33 WebGL2, r34 default, r34 WebGL2 |
| Parity and differences | M2b static, r34, r33 to r34 |
| Oracle and colliders | Oracle, r33 colliders, r34 colliders |
| Hub kit | Build, smoke log, attempt 1 log |
| Matrix | Update |
| M3 builds and hygiene | sizes, three import, revert diff, hygiene |
| M3 counted runs (B-09, B-10, B-13, B-08, contract suite) | final runs log |
| M3 units, types and lint | units, builder units, typecheck, lint, baseline |
| Tablet | state checks, tier, soak, device kit README |
| Look | contact sheet data, differences, D-18 pairs |
| X-04 attempt and hitch counts | notes, WebGL2 orbit counts |
| Hub kit (m3) | build, smoke, look runner, verify |
| Matrix (m3) | update |
| M4 builds, sizes and site resolution | build log, sizes, site resolution, its log |
| M4 hitches, look A/B and warm-pass concurrency | README, attribution, look A/B, concurrency A/B |
| M4 X-04 (first stable frame; indicative) | cold, warm, log |
| M4 X-11 and X-12 | bytes, allocation, allocation sites, log |
| M4 counted runs (contract suite, B-08, B-10, B-13, B-09) | final runs log |
| M4 gates | gates, full units, builder units, typecheck, lint, hygiene, licences, final gates |
| Hub kit (m4) | build, verify, smoke, verify after the smoke |
| Tablet (M4) | runs log, state, tier, soak |
| Matrix (m4) | update |
| M4 audit | final audit |
Site-owner handoff
The kit API is released as of KIT-M1: READY 2026-09-29; Golden Gate can consume it. The Farm rewrite matches the sealed pilot as recorded under D-20 on r33 and r34, carries the M3 features (mobile play under D-22, tiering, public UI, accessibility, pause and reduced motion) and the M4 hardening (the warm pass before ready, the WebGL2 hitch fix, the governor floor, this tablet class on minimal), and passes every counted local check on the m4 outputs, but it is not yet ready to replace the pilot on the website: the coordinator's hub runs of the m4 kit (X-01 on m4, X-04 as the first stable frame, X-05) and the decision on the hitch rule are pending, and the site's three alias, island check and B-14's Farm consumer belong to the site round. Keep the pilot fallback and current page behavior. No website, pilot, mirror or Kiln engine edit is part of this handoff.
The site links packages/farm as source, so its next build picks up the D-22 HUD, the fix for the 390 px overlap its island check found, and the M4 warm pass (the scene now compiles every pipeline before it declares ready, behind the loading screen, D-23); re-run that check there. packages/farm/src/index.ts exports are unchanged. Add the kit's three alias to the site's Vite configuration (the kit's sceneSourceConfig carries it): without it three.module.js joins three/webgpu and the m4 Farm chunk grows by 353,681 B / 84,334 B gzip, to 111.72 % / 108.02 % of its D-15 ceiling (evidence; m3: evidence).
The kit README and source consumer demonstrate sceneSourceConfig('public'): merge the source aliases, bare-three Clock compatibility facade, deduplication and compile-time flags into the host’s Vite configuration. The kit stylesheet is injected, so no separate CSS setup is needed. The host owns loading/error presentation, Exit and fullscreen; mount the scene as a client-only island.
Website infrastructure exception (C-03): preserve React DOM’s one shared nonpassive document selectionchange listener for the page lifetime. Scene teardown releases every scene-owned listener. Its count remains one through cold mount and all cycles; do not intercept document methods, remove the shared listener or patch React. Exact pinned source locations, cold proof.
Farm r33/r34 staging and exact upload file hashes are under packages/farm/staged/r33/SHA256SUMS and packages/farm/staged/r34/SHA256SUMS; all model, data and licence bytes verify, and r34 matches the current website model manifest. D-12 reserves uploads and CORS headers for the owner. D-11 reserves website mobile-copy changes for the coordinator after B-09 passes; B-09 now passes, emulated and on the tablet with real touch, so the coordinator may make them. D-18 keeps the Farm on ACES Filmic at .95; the ACES/Neutral pairs and the look contact sheet are ready for the owner's review. No publishing or deployment is authorized by this report.