Boids primary allocation and window GDI ownership repair
The two natural composition conflicts reproduced by f0e8a663c are materially repaired. A stale GPU-derived primary DIB no longer becomes a signature-polled CPU frame. Later real window GDI uploads compose their actual dirty rectangles over the accepted GPU front; a subsequent native GPU PRESENT replaces those older rectangles.
Native ownership identifies the live primary entry, its current DIB allocation, creator, destination window, RT and actual submitted PRESENT record. The owner is published before the ring record becomes visible. GPU Flip follows the swapped allocation. CPU readback stays GPU-derived through conversion and after read-only materialization; later watched CPU writes change the signature and regain normal CPU eligibility. Final release clears ownership before address reuse. This replaces the previous process-wide parked-slot suppression with checks on the selected allocation.
The compositor snapshots only the bytes in real window-DC upload rectangles. Disjoint patches retain their native PRESENT cutoff and instance identity, including delayed bitmap delivery. Exact instance/window/source checks and existing retirement guards refuse foreign or stale frames. Lost WebGL contexts fail through the existing stream error path before silently publishing a blank frame. There are no guest input, timing, stepping, pause, repaint or flush changes.
Qualification is in scratch/runs/20261117-boids-primary-gdi-ownership-repair/result.json. Authored native alternating GPU/GDI, actual full WM_PAINT and partial WM_PAINT sequences compare all 307,200 source pixels against independent literal blue/red/black/green or dirty-rectangle expectations. Native Flip tests cover read-only materialization, later public Lock/read/write, a retained two-pixel CPU write, actual previous-RT refusal, cooperative window switching, primary retirement and same-address client-allocation reuse. Shared native cooperative routing, surface-fence, actual Worker cancellation/retirement, software stream, DirectDraw placement and deferred GDI regressions pass. Optional headless-GL tests skip for missing native dependencies; actual browser renderWorker WebGL1 qualification supplies the GPU-path evidence.
A real renderWorker WEBGL_lose_context fault causes the next natural native record to fail with gpu-stream-webgl: context lost; no later frame is accepted. Expected loader-retirement failure logs remain recorded: this proves publication refusal, not successful guest shutdown under context loss. Browser and independent supervisor/watchdog processes are actually awaited.
The canonical locked build also found three stale generated SDK model sizes introduced by 35cc4b73b. Only those byte-count entries were corrected to match the existing local original media. All 38 distinct registered executable/media paths match their retained hashes; the earlier “152” count was four sets of 38, not 152 distinct files. No EXE, model bytes or application defaults changed.
Repair commit 755ec6578 was published by strict pull/push on main 7c84a0ae0. A subsequent Git-only native full WM_PAINT run on that published main reproduces all four expected literal states, with no unexpected post-GPU publication and all nine ownership/fault controls refusing. Selected complete partial and full frames also pass independent source/output literal oracles: 307,200 source pixels and 1,185,280 scaled output pixels per frame, with zero mismatches.
Exactly one fresh passive original ran on immutable published 7c84a0ae08088267e63a1b152173f672b65cc18c, module SHA-256 a105f6ce02932b3edebe9376a8b5eb4658f2159c52bb3cc76ae5ac4eadb05b0c. Debugger-observed module bytes match on page, guest Worker and render Worker. Readiness identifies actual guest/render Workers, WebGL1 and ANGLE/Intel Mesa. Public executor counters include 67 draws and one PRESENT at readiness; these are not evidence of visible motion.
The first complete-display request, after the fixed passive wait, exceeded the unchanged 3000 ms controller deadline and terminated as DISPLAY_UNKNOWN. There are zero complete original captures. The endpoint was closed without replay, rebind, further input, configuration/EXE/clock changes or another original. Consequently sustained original flock motion is unproved. The concrete next boundary is complete-display capture completion under this original workload, using retained evidence or a separately authorized follow-up; this result does not identify why the request missed its bound.
Actual runtime exit, browser close and both watchdog waits are zero, stdout/stderr reach EOF and no token-owned descendants remain. This is host lifecycle evidence, not a native guest Quit claim. No box was created (ID/expiry null, state NOT_CREATED, stop receipt N/A). Evidence uses Git commits and binary patches, existing media paths and hashes, two named selected authored complete screenshots and tested modules, with no source archive or capture bundle.