Boids GPU/GDI publication ownership, 2026-11-17
Diagnostic only: no production repair or sustained original motion is claimed.
The incoming GPU retirement/compositor changes do not resolve the authored
ownership witness. The same sequence reproduces on reconciled, published main
9025b309fa5b68ee5f48cd07cc6fb3fbae03fdc4, with native module SHA-256
33ef36c2de1c86ca1302b5659f02bb48f1794dfc4e20945821251198f04cb03e.
Evidence is scratch/runs/20261117-boids-gpu-gdi-publication-owner/result.json.
The new PE extends the retained Device2 RGB16/D16 indexed-draw fixture. Real clocks alternate GPU blue, GDI red, GPU black and GDI green every two seconds. A second variant paints through an actual registered guest WndProc, InvalidateRect, PeekMessage/DispatchMessage and BeginPaint/EndPaint. It never uses UpdateWindow or forces a host flush, repaint, idle, guest step, input or pause change. Original EXE bytes are not modified and no original is launched.
At completed ordinary publication the observer records exact window/instance, canonical presentation identity/version, GDI and GPU write sequences, selected layer identity, bitmap closure and the actual complete source pixels. Write call stacks distinguish guest GDI uploads from DirectDraw signature polling. The full-pixel oracle compares every one of 307,200 source pixels with literal blue/black rectangles, solid red/green, and white-clear references. It explicitly distinguishes the retained half-pixel raster convention from the incoming integer convention; the upstream subpixel change is not called an ownership failure. The two selected complete captures also pass all 1,185,280 output pixels, including scaling and letterboxing. They show white, not a repaired scene. Pixel agreement alone does not establish that the right writer won.
Two natural divergences are now joined within the same completed composition:
- In
reconciled-sequence, publication 8 selects GPU blue at layer write 3 while the window's canonical presentation version 3 contains solid red at GDI write 4. The accepted GPU layer reports non-stale. The WM_PAINT variant independently reproduces this at publication 10: red version 4/write 7, older blue GPU write 6, same window 65537/process 1000/instance. All source pixels match blue rather than the later authored red write. - In
incoming-causal, publication 10 shows a completely white DIB layer at write 8 after an accepted blue GPU frame at write 7. No intervening GDI upload replaced that frame: GDI remains write 1. The write-8 stack is_presentDxSurfaceToMainWindow→presentIfChanged→_presentBestDxOffscreen→WineAssembly._presentDxIfDirty. Reconciled main and its WM_PAINT variant retain the same white/old-blue DIB fallback outcomes between valid blue/black GPU frames. These are actual composed pixels, not an inference from draw counters or opaque frame receipts.
The next production boundary is the exact relationship between a GPU-owned
DirectDraw primary's DIB and window-DC GDI writes. In lib/host-imports.js,
_presentBestDxOffscreen checks the legacy parked-slot export, then signature
polls a primary DIB even when GPU-stream Flip owns its current pixels. In
src/09abg-d3dim-gps.wat, Flip swaps the DIB identities and presents the GPU
target; the fallback needs authoritative ownership for that exact surface and
allocation lifetime, including CPU materialization and real later CPU writes.
Separately, lib/renderer.js::_overlayFrameLayer does not retire/merge a D3DIM
GPU layer on a later window-GDI upload; its D3DIM stale closure tracks the CPU
DirectDraw present sequence rather than that GDI write. A blanket sequence test
or blanket suppression of DIB fallback would not prove partial-GDI composition,
retained CPU Lock writes, same-address reuse or source retirement correct.
No such partial fix was installed.
The original's earlier changing opaque GPU frames remain valid evidence from the Device2 localization run. These authored divergences do not prove which writer supplied its retained Paused screenshots. The optional passive-original diagnostic and conditional repaired-original run remain unconsumed in this task; no missing fact required spending them to establish this boundary.
All five authored browser runs used the actual guest Worker/renderWorker WebGL1 route and reported ANGLE/Intel Mesa, despite the requested launch flags. They completed normal browser shutdown, actual supervisor/watchdog waits of zero and no remaining owned runtime processes. Nine native capture-owner refusal controls pass per run. They do not constitute the requested future repair's complete native same-address/source-mismatch/context-loss suite. The focused observer/oracle tests, 23 capture/decoder regressions, lit-alpha regression and Node doorbell same-address retirement test pass. The last is notification-lifetime coverage only. The canonical flock build passes with an explicit subprocess wait; its shared-tree module differs from the immutable Git-only native module because foreign overlays remain uncommitted. The earlier build log completed, but its tool wait returned 143 and is not the qualifying build. A mistyped alpha-test path failed before the correct test passed.
No box, worker, source archive or public deployment was created. Source recovery uses exact Git commits, binary diffs, new-file patches and observer suffixes; media is referenced by existing path and SHA-256. Selected images and all actual waits are retained. Publication details and remaining limitations are recorded in result.json. The worker releases its owned diagnostic files after publication.