NFS III capture composition and read ordering, 2026-11-18

Evidence: scratch/runs/20261118-nfs3-capture-composition-order/. All predecessor UNKNOWN scopes remain closed.

The exact authored 64-draw Device2 control uses frozen c9 browser sources, published host-step lifetime corrections, the b8c8 repaired module and normal guest/render Workers. Its complete blue rectangle and white background were personally reviewed. The actual connected logical canvas and retained native layer have the same whole-pixel fingerprint (328607173), 76,800 nonwhite pixels and 230,400 white pixels. Four actual owner module readbacks match b8c8.

The new read-only helper records wall and absolute monotonic clocks immediately before and after every synchronous canvas read and every guard evaluation. In the first native comparison, the connected logical read took 113.75 ms, the presentation read 104.25 ms, the backing read 28.76 ms and retained-layer read 20.82 ms. The initial guard was ready with a 50 ms-old producer. After the synchronous reads the unchanged guard correctly refused the same producer at age 320 ms. This is a measured observer-ordering failure, not evidence that the screenshot itself returned with stale native context.

An ordinary timer turn permits queued metadata delivery within the original fixed three-second operation. The first reevaluation still refused the old producer. Ten milliseconds later a genuinely newer producer was 11 ms old, and the existing guard passed with identical scene, process, thread, window generation and target references. No old timestamp was renewed. No guest poll, wake, clock write, focus change, input or graphics mutation was added.

The helper pins the owner, native endpoint, Worker, window object, instance, process, renderer, source and output canvases, and render endpoint. It checks the selected top-level target and native layer again after delivery. Regression cases reject stale producers until the fixed deadline, retirement, endpoint or Worker replacement, reused HWND objects, foreign process/instance, source replacement, generation change, hiding and a target absent from the selected composition. Physical input retains its separate existing guard, immutable in-flight deadline and independently prearmed UP.

This corrects capture preparation only. It does not repair the remaining display-publication discrepancy or prove driving or game input consequences.

The first fresh page attempt stopped before registering or launching the original: clicking the hidden debug-toolbar Launch button failed. Its actual native/watchdog waits and browser EOF completed. The supported production agent launch action in lib/agent-remote.js uses the registry select followed by the public launchApp() function; it does not click that hidden button. An authored heavy native control reproduced the hidden-button refusal and passed through that supported route, with the same native module, scene, composition pixels and normal Workers. Its normally selectable Notepad catalog entry carries only the authored executable; Rock & Roll is absent from the normal selector, as a separate prelaunch negative confirmed. Neither prelaunch failure registered an original or sent physical game input. The corrected driver follows the existing public shell launch route without changing shell/runtime code, guest state, focus or scheduling.

After the helper (6423f80a4) and tested launch correction (5592ad310) were published, one original was actually registered and launched. It retained the genuine chooser, all 135 verified media files, frozen c9 source plus recorded published patches and b8c8 module. All eight live owner modules were read back as b8c8. The whole initial Corvette loading artwork was personally reviewed. Logical canvas, compositor source and retained layer had equal whole-pixel fingerprints. Initial diagnostic reads made the old native producer 432 ms old; the unchanged guard refused it. Ordinary delivery supplied a new native producer, accepted at age 14 ms with the same owner and generation.

At 236 presents the one post-loading screenshot took 370 ms and was again white beneath the host strip. Exact PNG CRC/decode counts 1,143,910 white pixels out of 1,228,800; its SHA-256 is the same 3c8f82d3 as the predecessor white page. A whole lossless re-encode was also retained for review. Preview transport remains unreliable; pixel counts and live canvas readbacks are independent evidence, not a MIME-based inference.

The diagnostic now isolates where those pixels differ. Both the connected 640x480 logical canvas and exclusive compositor source contain 307,200 opaque white pixels, zero nonwhite pixels, with matching fingerprint 2596822469. The connected visible 1280x926 presentation contains 1,143,610 white pixels. The retained 640x480 GPU layer instead contains 307,178 nonwhite pixels and 22 white pixels, fingerprint 200337884. The selected top-level window is 98305 in process 1000 with the actual instance; it is the presenter's exact window, its current layer remains selected, RT is 3 and layer sequence 21530. The recorded display backend is canvas-nearest, using the exclusive source and connected presentation output. Publication is 113, guest slice depth is 1, GDI depth 0, repaint is scheduled and Worker repaint is deferred. These observations identify a displayed-source/retained-layer discrepancy while publication is pending; they do not establish the causal producer or justify changing the compositor's publication fences. The earlier native association observation had matching DIB, creator, positive window/RT sequence and CPU stamp 1; no separate native DIB table read was taken at the final capture.

The guard immediately before diagnostic reads was ready with a 44 ms-old producer. After the reads it correctly refused that producer at age 294 ms. After ordinary delivery it accepted a new producer at age 34 ms, with the same selected scene and owner. One retained receipt had 318 ms transport delay through readback; it was not relabeled fresh. The 250 ms producer fence and exact target checks remain intact. Thus admissible native context alone does not certify that the current display contains the retained GPU frame.

The white image supplies no reviewable cockpit. The scope was closed UNKNOWN without binding or sending any physical input, and with zero later owner endpoint requests. No unchanged retry followed. Throttle, steering, Pause, Continue and ordinary game Quit remain unverified; browser cleanup is not Quit. No speculative graphics repair was made.

Focused ordering/identity negatives, passive scene preparation and existing native input/deadline/release regressions passed, and the locked full build passed. The final helper control and supported-shell control retained exact module readbacks and byte-identical authored pixels. Six supervised scopes completed six actual native waits and six actual watchdog waits, all with both EOF streams, leaked=false and pollError=null. The two nonzero native exits were the documented prelaunch negatives. The actual original retained 900/920-second custody and 180-second Quit reserve. No box was created: ID/state/expiry/stop receipt are null/NOT_CREATED/null/null. No external loss of this task's owned evidence was found.