NFS3 complete closure loads the game and reaches D3DIM; race qualification blocked by capture predicate

The corrected native route reaches the actual game on the production 7ec49 module. Native catalogue → setup → reload → setup selection → warning OK → setup OK → reload → chosen-game launch preserves D3D registry values and produces an active owning D3DIM render Worker. Acceleration, steering and road-relative displacement remain untested: a run-local capture predicate rejected the numeric renderer endpoint ID and terminated this route. This is a harness failure, not a demonstrated emulator mismatch. No emulator code changed.

Evidence is in scratch/runs/20261108-nfs3-complete-closure-race/. This continues native chooser qualification. Before allocation, the released test-staged-native-route.cjs passed all three production registrations over staged HTTP with size/SHA checks, including the separate nfs3demo.exe registration (1,577,984 bytes, SHA-256 0defab3eeb22ee4b6e0007a4d5b26a99d868008ba77e2b9bd3ef770e924548ad). The original media were referenced locally, not copied into evidence. All pinned inputs were verified locally and the remote process later verified 726 unchanged pins. runtime-pins.json describes execution; prepared-runtime-pins.json is explicitly a subsequent, unexecuted continuation.

Native consumption and actual renderer

retrieved/webgl/native-selection.json retains ordinary native setup writes: Direct3D Device 1, Thrash Driver d3d, Group D3D, D3D Device 0, Hardware Acceleration 1, Triple Buffer 0. Setup terminates before the page reload. No registry injection or renderer override was used. Five independently released mouse clicks occurred; one warning click used viewport-normalized Y instead of canvas-normalized Y and landed above its button. A fresh capture was reviewed, then the correct visible warning OK was clicked. This was a known successful command with the wrong point, not a transport UNKNOWN or input replay.

early-game-load.json requires exactly the chosen registered app, running, not stopped, with a guest Worker before input-chain creation. Its regression rejects seven missing/failed/wrong-app states and accepts the live game without inferring renderer readiness. The game EXE serves HTTP 200 with the expected SHA at 06:06:51.915Z. The initial reviewed game image is only the startup window; it does not show a race or establish playability.

The subsequent game-settled-state.json, saved before the failing screenshot binding, contains actual D3DIM activity:

Evidence Observed value
Registered app / process nfs3_chosen_renderer / 1000, running
Owning renderer target F5F103926047DB828414DABE96907CD3
Render endpoint numeric 1, API legacy, backend webgl
Owning observed context WebGL 2, 640×480
Context renderer ANGLE, Intel Mesa UHD Graphics 620 (KBL GT2), Mesa 23.2.1
Actual guest and renderer module SHA-256 7ec49cbbae4cb177909fd8fd9d3599713ee5f2a5ffb7f7e8c890848a540052bf
Draw calls / triangles 805 / 1766
Reported fallbacks / errors 0 / 0
Software / fallback software draw completions 0 / 0
Owning control ready 1, error 0, completed 4230, commands 1000

The context is observed in the same actual render Worker that owns endpoint 1; its module bytes were recovered through CDP. The adapter wrapper itself has no gl member, so that field is null; context identity comes from the Worker's passive context observer. This establishes active D3DIM following native D3D selection, not a measured FPS or parity result. The mounted d3da.dll HTTP hash is recorded, but HTTP delivery alone does not prove its guest module load address. No guest module-load inventory was collected before termination.

Precise first new dependency

At 06:07:05.608Z, capture() skipped establishReady() because my added predicate required typeof renderEndpoint === 'string'. The actual endpoint is numeric 1. The original qualified native chain therefore entered its capture path after renderer attachment and correctly asserted native stage requires no renderer: 1 !== 'none'. The pending capture command received the sticky terminal first fault; no game input or failed command was replayed. Seven earlier commands had exact successful ACKs. No acceleration or steering input occurred.

renderer-attached.cjs and its focused regression now refuse absent/null/none endpoints and accept numeric and string actual IDs. browser-run-prepared.cjs uses that predicate while preserving all owner/module/freshness/readiness checks. These are prepared only; the executed browser helper and pins remain untouched. The next fresh run must use this prepared correction, qualify renderer binding, review actual game images, then use ordinary menu/race input to test acceleration, steering and road-relative displacement. Do not resume this stopped box.

Validation and lifecycle

analysis.json verifies 1,259 complete pinned HTTP responses with zero mismatches; 16 PNG references pass CRC and zlib decoding. Original setup/selection/startup images were personally reviewed. All 130 remote artifact references (120 unique hashes) were retrieved and independently re-read for size/SHA before stop. Chrome exited zero; controller/Chrome PIDs, profile, writer lock, both ports and profile-associated processes were absent. Both independent supervisor children exited zero and the supervisor has no children. The messageboard watcher stopped. Chrome was 151.0.7922.108 with native 945×889 viewport and real guest Workers.

One UUID-idempotent box bx_9mp55buk was created after observable independent 25-minute watchdog arming. Initial expiry was 2026-11-08T06:29:47.686Z. Final state is ARCHIVED, expiry null. Normal stop receipt stop_5df2850db5433feb completed at 06:07:38.629Z, forced=false, error null. Evidence verification completed at 06:07:36.620Z before the stop request. Disk stayed above 5 GB. There were no guest patches, forced config/scene/clock/seed, private handlers, source/media archives, old-box resume/delete, RAM gate, deployment or workers. Git base, relevant binary diff, new-file patch and exact tested pins retain source recovery without a tree archive. No new build claim is made for this docs-only change. Full27, parity, audio and unmeasured FPS remain open.