Native OpenGL producer to real render worker under delay

The integration gap in the lease sweep is exercised by test/test-gl-native-render-load-web.js. The final run passed both WebGL and native software paths. No production defect was reproduced and no production source changed. This is a bounded integration result, not a general claim of crash freedom.

Evidence is retained in scratch/runs/20261102-gl-real-worker-load/ (ignored), especially evidence-review.json, retrieved/gl-load-final/observations.json, the two PNGs, final-locked-build.json, and RECOVERY.txt.

Actual components and fixture boundaries

The producer is a physical browser Worker instantiating the shipping Wasm over shared process memory. It calls the native encoder export, uses the native three-slot pool, and blocks through production GuestRpc.createWorkerImports. Production createMainBroker, GLRenderHost.Bridge, shared-packet pool, submission queue, RenderWorker.Manager, endpoint protocol, d3d-render-worker.js, gl-render-worker.js, and backend execute unchanged.

The harness supplies a minimal window/DC and real native-backed GDI surface, uses the production renderer's three canonical readback methods, and supplies the host's GL descriptor adapter. Unrelated host imports are inert. This does not boot host.js, an x86 application or a game. There is no FakeGL, VM worker double, semantic replacement consumer or replacement rendering result.

A worker wrapper queues real incoming render messages for 300 ms before delegating to the unmodified worker handler. Its observers hash submitted bytes and inspect contexts without replacing commands, replay or replies. The remote supervisor bounds the complete invocation to 90 seconds; the final invocation took 9.74 seconds.

Path Actual rendering/presentation Coverage limit
WebGL Actual worker GL backend; CDP SystemInfo reports ANGLE_SWIFTSHADER, SwANGLE 5.0.0 Software WebGL, no hardware GPU coverage
Native software Shipping WAT rasterizer in producer; actual render worker mirrors the leased software front and presents it Synchronous software lease; not the separate queued D3DIM draw executor
Context retirement Real backend deletion and owner invalidation, then a native producer present Cross-owner deletion record is deliberately fixture-authored shared memory, not a second native producer

The installed Chrome launcher adds a fingerprint seed. Its page-level WebGL identity reported Intel/Mesa, while browser GPU information identified SwiftShader. The fixture requires browser-level SwiftShader confirmation and retains both observations. Node was v24.22.1; installed Chrome reported 151.0.7922.108. Nothing was downloaded.

Independent output and ordering assertions

Each backend presents six frames with literal red, green, blue, yellow, magenta and cyan background pixels, plus a literal white packed triangle pixel in every frame. Native glReadPixels returns [0,255,255,255] and native GDI GetPixel returns COLORREF 0xffff00. Query and pixel destination canaries remain 0xa5a5a5a5.

Every native packet follows exact three-slot rotation. Hash/offset/length observations immediately before real consumption equal the published bytes, despite subsequent visible overwrites of reused slots. WebGL reports six leased frame batches, zero copied batches and zero copied bytes. The first two frame acknowledgements precede first consumption; the third waits 451 ms, until the oldest presentation retires. Live submissions stay at most two. All six frame results are 1 and their observed pixels retain FIFO order.

The WebGL queued finish waits 910 ms; native pre-GDI readback waits 613 ms; finish, backend RED_BITS query and glReadPixels each wait about 301–303 ms. RED_BITS writes literal 8 before the producer continues. Software frames each wait for presentation, and finish/readback wait about 301–302 ms. Software MAX_TEXTURE_UNITS writes literal 2 locally in the native producer; its 0.29 ms query is explicitly not evidence of a backend query barrier.

After cross-owner deletion both paths report zero contexts, invalidate active owners and return 0 from the native producer's subsequent present. RPC slots are idle. Endpoint closure removes ports, frame owners and layers. The actual worker acknowledges render-stopped with heapHead: 0; the manager invokes physical termination, pending requests drain, and CDP reports no surviving fixture worker targets. Browser Workers have no independent OS exit code.

First failures, validation and recovery

The first concrete unexpected result was the fixture's SwiftShader-only page-renderer assertion rejecting the fingerprinted Intel identity. Its six pixel frames and native output checks had passed. That run is preserved in first.tar.gz with its exact fixture source and input hashes. Two subsequent fixture failures are also preserved: unhandled production async dx_trace messages, then an incorrect software query expectation of 4. The harness now routes async calls through production serveCalls; current native source explicitly returns 2 for MAX_TEXTURE_UNITS. These were harness assumptions, not production fixes. No old-source production rejection experiment was appropriate because production remained unchanged.

Final Node PID 34247 exited 0; Chrome PID 34254 exited 0 without a signal. Earlier Node invocations exited 1 on assertions; each Chrome exited 0. None timed out. All four evidence archives and their 24 listed files matched SHA256 before shutdown; all 43 final served hash records matched pinned runtime files or recorded virtual fixture sources. The final archive SHA256 is 129f86001e302991dc6b7045a1fe14467ebe5b50dcb1636261b9fb9f3bd79e98.

The unchanged shipping module is 2,069,384 bytes, SHA256 6d9563ff4a1c089c9d0ac4153748411990dc3c92f3a50e2509bbe9ffaaca30a7. Both locked full builds passed. Final source verification matched 1,264 entries; all 1,263 entries apart from the new fixture also match admission. The original two accepted test files retain their hashes and assertions; neither was rerun or represented as integration.

Execution started only after DBTS19762 then Asghan25107 actual terminal, result/final, exact author release and helper-absence gates. Separate boat bx_3yvraw6k had finite TTL 5,400 seconds and original expiry 2026-11-02T17:51:41.686Z, recorded before execution. All named processes, crashpad helpers and temporary profiles were absent before the explicitly delegated stop. Retrieved evidence was verified first; boat state is STOPPED at 2026-11-02T16:27:48.027Z. Observed local free space stayed above the owner's 2,000,000,000-byte floor.

Recovery needs only a fresh finite-TTL box with installed Node/Chrome, the 17 pinned runtime files and this fixture. Run GL_LOAD_OUTPUT=/absolute/evidence/observations node test/test-gl-native-render-load-web.js from that minimal runtime, under the documented supervisor deadline. Detailed transfer, verification, failure archives and recovery commands remain in RECOVERY.txt; no stopped boat must be resumed to inspect the retained result.