TR2 native OK attempt — 2026-11-08

Native OK was not activated. This attempt failed because the worker repeated the image-review error explicitly identified in the handoff. No emulator defect or persistent Graphics absence is established. Current Venice controls and all presentation rates remain unqualified.

Evidence is in scratch/runs/20261108-tr2-native-ok-current-control/. No emulator code changed. Only this note is committed.

The released prior-run correction was checked against original-resolution images and hashes: startup, fresh-graphics and refresh-graphics are byte-identical and contain Graphics; only final native-after-idle differs. The same pattern occurred in this new run. Its startup.png, fresh-graphics.png and refresh-native.png are all SHA-256 964fd8396fb7b73def05796add7c09a02f01dc0f65feb1930d87f8315b8bf133. They contain the Graphics sheet and visible OK. The worker nevertheless described fresh/refresh as desktop-only in intermediate commentary. Those descriptions are invalid. This is an operator review failure, not evidence of an image decoder, capture tool, renderer or emulator defect. Final native-after-idle.png, SHA-256 e9bfc73f3e7e375dbb5479c85403403bb2e63f8906eb0a8f70da6e44b670691e, shows the desktop. Read-only metadata reports the dialog visible at publication sequence 5; its metadata does not settle why final composition differs.

The first three captures were at 05:55:34.495, 05:55:40.210 and 05:55:48.489 UTC; final idle was at 05:56:52.462. The 12-second admission window was not used to activate OK. No native guest mouse/key request was sent, no stale image was used for input, and there was no original replay or transport UNKNOWN. Four exact successful ACKs cover three passive captures and normal quit. Browser preflight independently delivered five balanced trusted key pairs on about:blank.

The actual reused mouse driver's local mocked-CDP test passes: normalized canvas point [0.6, 495/691] with canvas rectangle (10,160,605,691) produces native viewport (373,655), the visible OK center on these images. A pixel-valued point is rejected with zero send effects. Balanced independent release, strict owner/module/freshness checks and atomic ACK/sticky-UNKNOWN controls pass their local tests. These checks do not constitute native OK activation or gameplay.

The unchanged production route was index.html?debug&threads&app=tomb_raider_2_demo&gl-renderer=webgl&d3d-renderer=webgl. Published source commit was 957b861cd7c930f842f89504f87512ce36ea93f3. The actual guest Worker module is SHA-256 7ec49cbbae4cb177909fd8fd9d3599713ee5f2a5ffb7f7e8c890848a540052bf; no owning WebGL endpoint or context was created. The registered media and guest, config, clocks and seed were unchanged. Runtime pins, Git SHA, relevant worktree diff, helper patches and media path/hash references are retained; no source or media archive was made. PNG CRC/inflation, module identity, successful served response hashes and exact ACKs were checked locally.

No control acceptance or frame collector was reached. Persistent terrain-relative Up displacement, separate heading and weapon response are unproved on this run. Queue, rAF, publication and content-change counts were not reported as logical game FPS. Collector identity association, overhead, three warmed ten-second windows, median/p95 pacing, software comparison, full27, parity, audio and unmeasured FPS remain open. No repair or locked-build claim is made.

One fresh UUID-idempotent box, bx_crvykarg, was allocated after observable independent normal/watchdog delegates were prearmed. Its immutable 25-minute deadline was approximately 06:20:11 UTC; initial expiry was 06:20:11.191 UTC. One heavy runtime ran. Local disk samples remained above 6.3 GB; no RAM gate, resume, delete, deployment or nested worker occurred.

Chrome closed normally at 05:57:01.663 UTC, exit0/signal null. The immediate drain assertion retained a still-live controller PID; a later read-only check at 05:57:17.189 verified both PIDs absent, both ports refusing, no profile, writer or matching processes, and all 598 input pins unchanged. The initial retrieved full-drain.json preserves the failed early check; the settled check is retained in settled-drain-remote-result.json. A second inventory invocation was refused locally by its exclusive attempt filename before any remote action; the already retrieved evidence was independently verified before stopping.

All 62 original inventory references, 55 unique hashes, were retrieved and independently reread for size/SHA-256 at 05:57:17.381 UTC. Final box state is archived, active expiry null, stopped 05:57:18.610 UTC. Normal-stop receipt stop_cd90714028f1cf59 completed 05:57:18.630 UTC, forced=false, error=null. Both delegates exited normally and supervisor children are empty.

Next work needs immediate original-resolution and hash-correlated review of fresh visible Graphics followed by the validated normalized OK activation within the unchanged freshness deadline. The present run supplies no technical blocker to that activation; it supplies a documented operator failure. Controls must work before any presentation-rate qualification.