NFS III completed-frame publication lifetime, 2026-11-18

Evidence: scratch/runs/20261118-nfs3-deferred-publication-visible-frame/. All predecessor UNKNOWN scopes remain closed. No renderer fence has changed.

tools/nfs3-deferred-publication.cjs observes existing Presenter, slice, composition and source-staleness calls. It preserves synchronous return values, exceptions and receivers. It neither reads pixels nor calls guest exports, polls/wakes an actor, changes freshness, schedules work or paints. Optional native state reads use the tested module's region map and read the primary's published DIB/RT/creator/sequence, current allocation and CPU-present sequence. The sequence is checked twice; an incoherent record is explicitly identified. Its event repository and stale-wrapper restoration state are bounded. Retirement, HWND reuse, foreign instances/processes, hide, source replacement, GDI depth, active/poisoned writers and endpoint closure remain observations; this witness does not authorize input or mid-slice publication.

The final authored 64-draw Device2 control uses the recovered exact c9 browser graph plus published host-step corrections, b8c8 native module, a separate native USER actor and ordinary guest/render Workers. Five accepted completed frames were observed, three while publication was held, with five ordinary publications during the trace. The longest observed interval from a held completion to a later publication was 2718.90 ms. Actual compositor stale predicates all returned false. Native primary DIB/creator records matched. The whole blue rectangle and white background were personally reviewed; retained GPU, connected logical canvas and exclusive compositor source each have 76,800 nonwhite and 230,400 white pixels, fingerprint 328607173. Thus a long await with a pending repaint can recover through ordinary boundaries; that observation alone cannot explain the original's white page.

Earlier authored scopes are retained honestly: the first trace did not yet distinguish rejected deliveries; a second control failed an unnecessary minimum-frame-count assertion despite correct pixels, and a later startup rAF poll hit its immutable protocol timeout before witness installation. The driver now uses finite external startup observation and records accepted deliveries separately. No failed scope was reused. These are harness controls, not original-game retries or evidence of a rendering repair.

Focused witness tests cover direct native-table reads, allocation mismatch, uncommitted sequence, invalid entry, bounded retention, source replacement, restoration ownership and exception neutrality. Existing Worker GDI publication fences, capture-ordering and native freshness regressions pass. The locked full build passes. No causal runtime correction is claimed by this preparation.

The fresh original after preparation exposed an earlier, ordinary boundary failure: _overlayFrameLayer returned CPU layer objects 10–13 instead of retained GPU layer 5, with guest/GDI depths zero and native allocation/creator valid. Every actual GPU stale predicate returned false. Publication 17→18 and 19→24 published those replacements. The subsequent held slice merely preserved an already white source. The whole page was reviewed white and the scope closed with zero input; no endpoint was reopened.

The repair gives native GPU presents the same scan-out priority as explicit CPU primary presents. d3dim_gps_window_primary_slot(hwnd,dib) validates the actual submitted HWND, current entry/DIB/creator, CPU-present stamp and stable publication sequence. CPU fallback selects that exact primary before scoring unrelated offscreens. Existing derived-byte, parked/scroll hold and all publication fences remain. Actual retained CPU writes still upload; a later CPU present, changed cooperative HWND, old DIB or retirement revokes the GPU pin. This adds no paint inside a guest slice and changes no freshness clock.

The authored Device2 fixture uses ordinary native Lock/write/Unlock before GPU drawing, so the completed GPU frame has a black CPU backing while an unrelated offscreen is white. With the same EXE, separate USER/DirectInput threads, 64-draw pressure and normal browser/render Workers, old source publishes 307,200 opaque white pixels (fingerprint 2596822469), while retained GPU pixels remain the blue quarter (fingerprint 328607173). The new source, connected canvas and retained GPU all match that complete blue-quarter frame. Both whole screenshots were personally reviewed. The old diagnostic correctly rejects the replaced selected layer; its complete pixel report is retained.

The native producer/software-executor/host regression also fails when only the arbitration block is removed (CPU fallback returns 1 and steals the GPU layer), and passes with the repair, checking every retained GPU pixel. Earlier native comparison with no registered window did not exercise a destination; that was corrected by registering the native window table before comparison. Native negatives cover wrong DIB, HWND switch, CPU-present stamp, retained native writes, Lock/readback, retirement, reuse and queued Flip backpressure. 8/16/32-bit host tests cover whole pixels, real CPU replacement and unowned fallback. Existing GDI/guest publication fences and retained-primary tests pass; locked full build passes.

A separately coordinated, backwards-compatible entropy_u32 host import is included because the gears owner's in-progress native import caused an actual LinkError in the shared harness. It uses Web Crypto CSPRNG only, without a weak fallback; that owner tested real native/compat and normal browser Worker pointer semantics. No gears native files or foreign render Worker files were edited here.

Published cf7044e37 was exercised on a fresh exact original (135 media pins, eight actual owner module readbacks, b91a6cdc native module recovered from the b8c8 source plus this native patch). Whole Corvette loading and whole rainy cockpit at 0 MPH were personally reviewed. The actual DIB table now confirms slot 14 is the native 640×480 RGB565 primary with entirely black CPU bytes; slot 18 is an unrelated 640×480 offscreen with every byte FF. The compositor source and retained GPU match. This directly validates the causal repair on the original, not only the authored fixture.

The first requested throttle did not reach input preparation: the screenshot was 17,736 ms old when the command was processed, so the old driver passed a negative preparation budget. The scope closed UNKNOWN with zero physical edges and actual native/watch EOF. It was not reopened. The association commit label in that scope's environment was mistyped; a retained erratum records this. Exact recovered source, module bytes and eight actual owner readbacks are authoritative; the runtime was unaffected.

reviewedInputAdmission now refuses an expired/future screenshot or insufficient preparation reserve as known NOT_SENT before any browser request. It retains the original screenshot timestamp, unchanged 15s lifetime, 3s native preparation, 500ms native confirmation and independent 250ms physical UP. A fresh capture and its filesystem handoff are retrieved together to avoid spending the review window between requests. Focused tests reproduce the exact negative-budget failure and check known refusal, boundary, clock and admitted reserve cases.

Published admission preparation c2256a77a was exercised on another fresh registered exact original, with the same b91 native module, recovered c9 graph, 135 verified media pins and all eight actual live owner module readbacks. Its whole cockpit was reviewed during countdown and again after race start. The race-ready capture is valid and visibly shows timer 0:05.81 / 1 MPH creep before any input. That autonomous creep is not evidence of user driving.

The first throttle intent was correctly NOT_SENT: only 37ms of its original review lifetime remained. A fresh combined capture/review retrieval then exceeded its fixed three-second screenshot deadline. The scope closed UNKNOWN before guardedKey, native input preparation or physical DOWN. A precomposed filesystem key intent was written after that capture closed; the driver never read it (no command-5, input admission, operation, release child or physical edge). No later browser/owner endpoint request was made, and no scope replayed.

The concrete next blocker is bounded browser capture/review admission under actual load. Its timing needs authored qualification before another changed original run; this task does not extend or bypass the in-flight deadlines. Original input-controlled speed/road consequence, steering, Pause, Continue and ordinary Quit remain UNVERIFIED. The rendering repair itself is causally proved by authored native/browser oldfail-newpass and by the fresh original's actual DIB table, matching compositor/GPU pixels and visible full cockpit.

All thirteen supervised scopes have actual native and independent watchdog pipe-close waits (26 total), both stdout/stderr EOF, no leaked containment and no poll error. All original scopes are closed. No boxes were created: box ID, expiry and stop receipt are null; state NOT_CREATED. Evidence, recovery patches, actual module/media identity and selected named PNGs are retained in the run; result.json is written last. No foreign render Worker hunk was edited.