Boids Device2 pixel localization, 2026-11-17
This is a diagnostic publication, with no production repair. The next boundary is ordinary display publication ownership/order after acceptance of the GPU window layer. The original does produce changing, opaque scene pixels in its actual renderWorker. One complete-display capture visibly contains birds and terrain; the next two display Paused. Sustained visible flock motion is not established.
Evidence: scratch/runs/20261117-boids-device2-pixel-localization/result.json. Runtime source was frozen at Git 972125ca472c2cff0b457daa9425885fcdb5a9e7, not the concurrent working tree. Recovered module SHA-256 is 1360defeb5faa36d2007b4c4c3b5651e8c11e0a9fd80c4b07661b220e09c75b1. Launch patches reconstruct all five pinned controller files for the qualifying blue/black runs and the passive original. Each injected observer suffix, base/served hash, requested source hash and media hash is retained. No source archive or media copy was made.
The authored fixture reconstructs the retained SDK/public command path: actual Device2 ABI, exclusive 640×480×16 flip chain, attached D16 depth surface, Viewport2, viewport color/depth clear, indexed legacy LVERTEX rectangle, BeginScene/EndScene and public Flip. Public GetSurfaceDesc checks 16 bits; the fixture requests the ordinary RGB16 display format and uses zero diffuse alpha with blending disabled. An initial public GDI text write exercises coexistence with a GDI backing. It does not read private simulation state. The first attempted fixture incorrectly reused Device3 viewport slots and exited 19; it was corrected from the retained method table. That failure is not a production defect.
On the actual native guest/renderWorker WebGL1 backend, blue and black variants differ only in submitted diffuse color. Both pass an independent literal oracle over all 307,200 completed source pixels and all 1,185,280 scaled display pixels. The rectangle contains 76,800 pixels; the remaining source pixels are white, all opaque. The oracle accounts for the documented GL half-pixel edge convention (x=0…319, y=1…240). Completed GPU framebuffer, pre-transfer bitmap, received bitmap and accepted window layer have matching bounded pixel summaries within each observed PRESENT. Blue FNV-1a is 2589337029; black is 657411525. Each native run also passes nine capture/lifetime refusal controls. Actual renderer identity is ANGLE/Intel Mesa WebGL1, regardless of the requested browser launch flags.
This rejects zero diffuse alpha as a sufficient cause of invisibility on the tested opaque RGB16 path. It does not qualify an ARGB render target or zero-alpha background-clear variant; this fixture uses the default white viewport clear. The public descriptor assertion checks bit depth, not explicit RGB masks. Stored alpha and render state remain unchanged; no global alpha force was added. Existing lit-alpha and immutable capture regressions pass. Notification retirement controls pass; the optional headless-GL regression reports SKIP because native dependencies are absent, so it is not counted as context-loss coverage. A software backend control completed its capture and browser shutdown, but its supervisor wait receipt is missing following tool exit 143; it is not counted as a fully qualified lifecycle run or native GPU evidence.
The one authorized passive original was launched only after recording the missing fact: whether original completed GPU pixels survive transfer to the accepted layer. It used the pinned registered Boids manifest (37 file entries plus EXE), defaults, real clocks, no input, prearmed independent 240/260-second preparation and 300/320-second runtime bounds. This was diagnostic authorization, not the conditional repaired-original acceptance run. No production repair qualified that latter run, and no second original was launched.
For original endpoint 1, worker epoch 1, hwnd 65537/process 1000, PRESENT sequences 14756 and 14832 produce different full-frame summaries (1628007074 and 2126421961). Each summary survives completed framebuffer → bitmap before transfer → received bitmap → accepted layer, with write sequences 733/734. All 307,200 pixels are opaque. The scene includes dark blue background, terrain grid and colored geometry. This is direct original GPU evidence, beyond changing command bytes.
Ordinary compositor observations also show a scene layer copied into an exclusive destination with globalAlpha=1 and source-over. Those observations precede the selected GPU samples and are not falsely joined to their publications. Complete-display publications 132, 318 and 497 are immutable captures: the first visibly shows the scene; the latter two show Paused. Post-capture content-boundary reads can see a scene layer alongside a Paused composition, but occur later than capture completion and cannot identify which write supplied that captured publication. PaintPaused remains a multi-condition WM_PAINT fallback, not proof that the simulation is paused.
The next causal experiment should reproduce and join ordinary GPU acceptance, GDI backing writes, layer selection and exclusive destination publication by exact generation/write/publication identity in the authored equivalent. The earlier no-GDI fixture's white final composition and later current layer were time-separated observations, not proof of a dropped same-publication frame. No narrow production defect has yet been demonstrated, so no production file was claimed or changed. Another original run needs fresh authorization unless the owner's conditional material-repair requirements are actually satisfied.
Final focused results, canonical build status, publication hash and lifecycle receipts are in result.json. No sandbox/box was created; box ID/expiry are null and stop receipt is not applicable. The qualifying native runs and original completed normal browser shutdown with actual supervisor/watchdog waits of zero and no remaining runtime processes. Selected screenshots retain authored blue/black and original scene/Paused boundaries. The worker releases its files and stops after publication.