NFS III native scene latency, 2026-11-18

Evidence: scratch/runs/20261118-nfs3-reviewed-frame-native-scene-latency/. Earlier original/UNKNOWN scopes remain closed. One fresh original was launched after the tested runtime correction was published as 32e695ea4.

The authored Device2 control creates its native UI on a separate thread and uses normal guest Workers and a real software render Worker. The frozen c9 browser source, exact repaired b8c8 module, native wait source and Flip association repair are retained. Diagnostic transport tags record producer, send and page receipt clocks without renewing any sample. Bounded page long task records distinguish delivery delays from gaps between native producers.

At eight draws per frame, 103 received samples had a maximum 1.85 ms Worker queue delay and 60.89 ms transport delay, with no receipt over 250 ms old. The largest producer gap was 683 ms. Its screenshot took 89 ms. That PNG was subsequently removed externally before personal review; its pixels are not claimed as reviewed evidence.

At 64 draws per frame, the old runtime produced gaps as large as 4592.71 ms. The maximum queue delay was 0.35 ms and maximum transport delay 49.44 ms; none of its 41 retained receipts was late. Page long tasks did not explain those multi-second gaps. A native scene 362 ms old correctly refused binding. A separate passive preparation wait expired after three seconds. This is not a sufficient repair and did not authorize a screenshot or input.

The normal free-run scheduler stopped giving sibling Workers slices when 250 ms passed without a new host step, even while the current host step was still awaiting a long main-Worker slice. Rendering backpressure can hold that slice for seconds. A responsive page and live native UI therefore lost their ordinary scheduling chain until the next host step.

The correction gives that awaited parallel rendezvous an explicit activity token. The token pins the backend and checks the running host's pause, stop and frozen state. The host releases it in its existing finally block. Both normal free-run redispatch and satisfied-wait offers recognize the live token. Outside it, the old 250 ms stale-step cutoff remains unchanged. No timestamp is renewed, no native poll or wake is forced, and no guest clock, focus, window state, rendering or input guard is changed.

A final old-code replay of the authored control verified all four actual Worker module readbacks before applying pressure. At 20 presents its native scene was 267 ms old, readiness and binding were false, and the producer gap reached 8071.28 ms. The main slice was still pending while the old scheduler reported a 498.93 ms-old host step. Maximum queue and transport delays were 1.46 ms and 30.06 ms. No physical input was sent. Both comparison scopes completed actual native/watchdog waits with both output streams at EOF.

The corrected 64-draw control completed 20 presents with 408 native producer samples. The largest producer gap was 266.74 ms and transport delay 56.64 ms; none arrived late. The main actor's longest synchronous render kick lasted 5637.14 ms, while the separate UI continued its ordinary native waits. Its 151 ms whole-page screenshot ended with a 30 ms-old ready native scene. Actual guest and render Worker byte readbacks matched b8c8. The whole authored blue rectangle, white page and host strip were personally reviewed. The PNG hash equals the earlier authored baseline, 2542b8560f3c313664235302d608184480f0a03929cb8a0caa509421e2ec063a. These are control pixels, not game progress or proof of observer neutrality.

Focused lifetime tests cover ordinary redispatch, satisfied waits, pause, stop, backend replacement, completion, nested acquisition refusal and idempotent release. Capture preparation only observes the existing guard within three seconds and pins owner/generation references; its negatives include timeout, endpoint/Worker/window/process replacement and retirement. The existing native producer, input, physical release, generation and legacy graphics guard regressions remain applicable without weaker bounds.

Shared external deletion removed retained PNGs, WASM, EXE and authored C inputs during this task. Exact b8c8 was recovered from the recorded Git base and patches, first reproducing 7d347 and then verifying the repaired hash. The authored C was extracted from this worker's own persisted command output; recompilation reproduced the original 64-draw EXE hash exactly. Missing visual evidence was not recreated by reopening a closed runtime.

Focused scene-latency, host-step lifetime, 59 scheduler cases, native source attribution, legacy input and independent physical release checks passed. The locked full build passed. Evidence logs and exact source patches are retained in the run directory.

The fresh registered original used the recovered b8c8 module, frozen c9 browser graph plus recorded published patches, genuine chooser and all 135 verified media files. All eight live owner modules (seven guest Workers and one render Worker) were read back as b8c8. One earlier nonowner Worker had already retired; its bytes were not invented or inferred from the live ones. The whole initial Corvette loading artwork was personally reviewed and had a fresh native scene. No input was sent.

At 122 seconds after registration, passive preparation returned ready after one observation. The whole-page screenshot then took 527 ms. The subsequent native guard returned ready=false, closing the scope UNKNOWN immediately. No later endpoint requests or physical inputs occurred. This screenshot was personally reviewed: the game area is white, beneath the blue host strip. Offline PNG decode counts 1,143,910 white pixels of 1,228,800, and a lossless whole-image re-encode shows the same white area. This is a captured blank page, not evidence of a driving cockpit or a claim about the guest DIB. The selected render layer still had nonwhite pixels and exact positive window/RT association 78211, matching DIB and creator, and CPU sequence 1.

The retained final 512 native receipts had maximum queue/transport delays 0.35/34.17 ms and no receipt over 250 ms old. Samples continued at roughly 100 ms intervals through screenshot completion. The postcapture evaluation header was 89 ms after its recorded producer, but precedes its synchronous canvas getImageData and guard calculation. The following association header was 253 ms after that producer; the exact guard calculation time was not recorded. Consequently the final refusal must not be described as proof that the native sample was already stale when the screenshot returned. Capture presentation and diagnostic ordering remain unresolved, and the refusal is retained unchanged. There was no admissible reviewed driving frame, throttle, steering, Pause, Continue or ordinary Quit proof. No unchanged original retry was made.

Detached custody retained the same 900 s native deadline, 920 s independent cleanup and 180 s Quit reserve. The driver exited 1 after its guard refusal; the watchdog exited 0. Both actual waits and both EOF streams completed, the supervisor reported leaked=false and pollError=null, and the browser actual wait completed. No box was created: box ID/state/expiry/stop receipt are null/NOT_CREATED/null/null.