NFS3 native scene freshness, 2026-11-18

Evidence: scratch/runs/20261118-nfs3-native-scene-freshness-cause/. The preceding dynamic inventory original and its UNKNOWN scope stay closed.

The guard's old observedAt is page Date.now() at a brokered check_input. The Worker negotiates the native protocol once and samples actual check_input calls at 100ms intervals, using its own performance clock. Empty local input polls can bypass the broker. More significantly, an empty native GetMessage parks at yield 7. The normal Worker repeatedly executes has_pending_message and checks the published input queue, without invoking check_input again until work appears. Requiring a fresh check_input before allowing physical input therefore rejects an otherwise live, valid, waiting UI. This is a freshness source mismatch, not proof that the renderer starves runnable UI work.

An authored x86 control creates its window and GetMessage pump on thread 2, with separate rendering and DirectInput actors. It uses real normal Workers, the frozen c9 source and native 7d347 module, and software Device2 rendering. The light observer-on control completed 60 presents while its UI performed 8,164 native has_pending_message calls. Its last ten check_input calls were startup work; the native UI retained yield 7, the same process/thread, focus, window and generation. The old guard was healthy but stale and unready. Fresh light and eight-draw pressure controls, with attribution off and on, both reproduced this. Observers add work; sequential shared-host measurements do not establish neutrality or a speedup.

The correction samples native USER getters immediately after the existing has_pending_message probe, at most every 100ms and only after the browser native protocol was negotiated. It sends a reserved negative input_validate metadata token through the existing asynchronous transport. It never polls input, wakes or resumes a wait, changes guest clocks/state/memory, modifies rendering, or reports input consumption. It adds no native module variant or interpreter optimization. The producing actor supplies its wall time and performance.timeOrigin + performance.now(); receipt time cannot renew a late sample. Both ages must remain within the original 250ms fence. The exact live actor link, Worker and endpoint remain pinned; stopped, stopping, retired, replaced, delayed and future samples are refused.

Fresh corrected light and pressure controls completed 60 actual software presents with native_message_wait scenes ready under the unchanged fence. The light checkpoint was 21ms old. All six old/new light/pressure whole-page PNGs match exactly, SHA256 2542b8560f3c313664235302d608184480f0a03929cb8a0caa509421e2ec063a. The whole authored blue Device2 region, white page and host strip were reviewed; these are control pixels, not game effects. Actual dynamic module inventories verify native 7d347 bytes. The native getter test preserves return 127, exact imports and memory, distinguishes four native wait probes from one input poll, and covers synchronous/asynchronous instantiation. Focused guard tests reject late/future clocks, retired/reused actors and replaced Workers before and after binding. Legacy native input and GL guard regressions remain passing.

A separate authored input qualification closed before input when personal review did not arrive by its finite review deadline. It made no DOWN, no bind and no later runtime request. It is retained as a failed scope, not replayed or relabeled as an input success. Existing native USER receipt and independent UP proofs are reused only for their unchanged claims.

The corrected native controls and focused checks, plus the locked full build, precede publication and the authorized fresh registered original. Its frozen c9/7d347/135-media/genuine-chooser preparation added only the tested native wait source correction and the bounded attribution diagnostic; dynamic actual owner inventory, binder and sampled reader remain in use. The original 500ms confirmation, 15-second gameplay review, independent UP, operation/identity checks and 250ms native producer fence stay binding.

The correction is published as 5aa5dabdf; strict pull/push both returned 0. The first locked build encountered another owner's silent-handler inventory WIP; the final locked build passed after that owner corrected it. Native source/guard and legacy input checks passed before publication. The existing GL native guard regression also passed. One invocation of the separate GL browser test omitted its required arguments and failed at argument validation; it is not an executed GL native case or a claimed pass.

The one fresh original registered and closed normally under independent finite and disk watchdogs. Dynamic inventory recovered the actual seven guest modules and one render module with 7d347 bytes. Retired slot 2 / endpoint 3 retained its normal-return and thread-ID reuse provenance with unknown byte identity null. The actual UI owner, thread 2 / slot 1, made 57,812 native has_pending_message calls but only 20 startup check_input calls; the final input poll preceded the last wait probe by over five minutes. Its native scene remained at GetMessage yield 7 / PC 0x4cba8e, window/focus 98305, process 1000, generation 1. Native producer samples now refresh the guard. This confirms the repaired source in the game, without claiming that idle wait probes consume physical input.

The main actor made 1,346 GPU stream kicks taking 293,837.61ms in aggregate, maximum 2,439.08ms. These durations include synchronous waits and are not CPU time. Entry/exit records distinguish synchronous mode 1 from mode 0. The UI's largest sampled between-run interval was 2,472.42ms, but an idle native wait does not establish starvation of runnable work. Neither a wrong guest clock, an interpreter defect nor a runnable starvation branch was proved. Attribution and readbacks add observer work; no neutrality claim follows from the controls.

Whole-page reviews found an actual cockpit, road, green opponent and rain, with 0 MPH, race time 0:01.67 and 1/1 laps in original/cockpit-review.png. Its native scene was already 1,229ms old at capture, so the unchanged fence correctly denied readiness. Other reviewed checkpoints were white below the host strip, including one with a 15ms-old ready native scene. None supplied both a fresh reviewed driving frame and admissible native scene. No bind, DOWN or UP was sent; throttle, steering, Pause and ordinary game Quit remain unproved. Process cleanup is not ordinary Quit. The unused admission helper was never invoked. Closed scopes were not replayed.

Read-only witnesses on that same original isolated a concrete remaining composition condition. At 272 presents, the retained 640x480 GPU layer had nonwhite sampled pixels, but native d3dim_gps_window_present returned zero for both the window lookup and render-target 3 lookup. The presenter's stamp and CPU present sequence both equaled 1; window and instance identities matched, and merged was false. Nevertheless layer.stale() was true and _overlayFrameLayer selected no layer. The visible connected presentation canvas sampled white, agreeing with the white whole-page captures. In frozen lib/host-imports.js the zero native window-present association is itself a stale-layer predicate. Its producer/retirement cause is not yet isolated: these measurements do not justify removing the fence or changing composition ordering, pixels or memory semantics. No speculative rendering fix or extra original was attempted.

All seven supervised scopes have actual native and independent watchdog waits with both stdout/stderr EOF and no leaked descendants. The authored review-timeout scope is the sole expected nonzero native exit. Evidence retains Git recovery and owned patches, unique authored inputs, module/media recovery identities, measurements and three selected whole-page PNGs; exact duplicate captures are recoverable by hash mapping. No box was created: ID, expiry and stop receipt are null, state NOT_CREATED. The result closes with the freshness repair tested and published, and the native window-present association as the remaining concrete blocker.