D3D software-worker isolation and refused-draw sweep — 2026-11-01

All three original standalone synthetic fixtures passed sequentially. No unexpected native crash, unhandled Wasm trap, assertion failure, signal or timeout occurred. This establishes the specific software-worker properties below; it does not establish gameplay, full-27 coverage or general crash freedom.

Evidence run: scratch/runs/20261101-d3d-worker-failure-crash-sweep/ (ignored). No production source, test assertions, shared helpers or compiler files changed.

Source and executable identity

Requested baseline: 5e46345eabbdfc5e88d03b13f9058715fd4efc85. Tested HEAD: 896edb18e9eff1586fd5b2320a322335ba9ba8c1; its only committed changes from baseline were TODOS.md and ops/STATUS.md. All 4,462 recorded inputs matched the accepted build's inputs-after.json before execution and remained identical afterwards. Tracked input bytes matched baseline. The 67 distinct loaded repository/installed modules were hashed at process exit and checked against current bytes; all 45 tracked loaded modules matched baseline. The other 22 are installed dependencies, recorded by exact hash in evidence.

File SHA-256
test/test-d3d9-software-worker.js 3c7c121647983a6e16448bbaec145347e3fe1729b56cee7a52d054b6e4a0efb7
test/test-d3d9-worker-survives-draw-failure.js 4af416d14a29ccdbff439d62fb8578250220f99c5769b0dca8586c25264e8227
test/test-d3d9-worker-alive-after-refused-draw.js 1e342c6fe39d11179803841e3de6064f67d6cead323fb55a526a2dc06479e19a
lib/d3d-render-worker.js f3d77c20eb14a8f72233216af03d2a50e60862b2fd10202e4970878050124b52
lib/d3d-command-stream.js ea1080750e8557f261d2b62f646b2cbc317be1b639cdc566b022c9a6ad494945
lib/d3d9-software-backend.js d73d855a6936881712c7b75e203cfc717c43dd82ca38fe67a576f4fc37da93ee

Both native fixtures compiled the canonical source closure. The two captured modules are each 2,024,992 bytes, SHA-256 ac9beb2ae3f746c621d0f258d4d3a2d1910207b0006e13ed96856bd0ac4975d9. Each has all nine non-custom sections byte-identical to the accepted primary build, including code and data. The primary is 2,025,029 bytes, SHA-256 2be3cebaa1c08a8e12075de000bfb83fb8692841c1452d9e317203f5c2fce8cf; the difference is its 37-byte custom stamp. Workers instantiate the compiled module passed by their parent over shared memory. The receiver-only fixture does not compile or execute Wasm.

Reused accepted flock /tmp/wine-build.lock bash tools/build.sh PASS evidence from scratch/runs/20261101-d3d-typed-stateblock-injection/build.log and outcomes.json (46.479 s). Inputs and primary bytes are unchanged; neither that build nor previously accepted tests were rerun. Runtime: Linux x86-64, Node 24.18.1, V8 13.6.233.17-node.50; no WINE/NODE environment overrides.

Commands and results

For each row, in this exact order, the command was timeout 120s node --require /home/user/wine-assembly/scratch/runs/20261101-d3d-worker-failure-crash-sweep/capture.cjs test/FILE. The evidence-only preload records Wasm bytes and loaded module hashes; it delegates compilation/instantiation unchanged and does not alter assertions, executors, source transforms or responses. SWEEP_CAPTURE_PREFIX selected the corresponding evidence filename. Exact argument vectors are in outcomes.json; run.py reproduces the sequence and stops on unexpected native trap, signal or timeout.

FILE Exit Elapsed Result
test-d3d9-software-worker.js 0 4.247 s PASS
test-d3d9-worker-survives-draw-failure.js 0 0.040 s PASS
test-d3d9-worker-alive-after-refused-draw.js 0 3.917 s PASS

What the passing assertions cover

The software-worker fixture uses three actual worker_threads instances and production WAT software rasterization. It checks direct-versus-worker full pixel equality and a known BGRA pixel, immutable input snapshots/readbacks, consumed-versus-completed ordering, parent allocation sentinels and x86 EIP/ESP preservation. Native query counts cover overlapping/restarted/released queries, wireframe, point fill, point-list clipping/size/batch boundaries, user clip planes and stencil rejection. Reset retires query brackets, changes target dimensions and preserves independent A/B depth contents. Graceful cancellation retires command leases only after exit, reports zero remaining native device allocation and proves parent reuse of handed-off heap storage. Intentional forced termination releases the transport lease only after exit and reports ORPHANED; it does not prove killed-worker heap reclamation.

The receiver-only fixture schedules JavaScript executor throws. It checks healthy ordering, per-command consumed/failed replies after a refused draw, suppression of subsequent execution on that failed stream, independence of a second device, recovery on a higher generation, and continued rejection of bad protocol versions or unordered descriptors. This is mocked receiver failure coverage, not evidence of a live thread or native pixels.

The final fixture starts one actual render worker with two software devices. A clip-space vertex with positive w = 1e-39 causes native projection/raster execution to refuse the draw. A subsequent draw on that device receives earlier render command failed; after 250 ms the worker thread remains alive. The second device then completes a valid native draw with a clean queue, and the worker is still alive. This fixture does not read back pixels or exercise real-worker generation recovery; receiver recovery is covered only by the preceding fixture. It loads test/binaries/calc.exe solely to establish an image base and executes no guest instructions. Final termination is the fixture's intentional cleanup, not an unexpected crash.

There is no real GPU or browser coverage here. BlackWhite2, held guests, replays/diagnostics, REP work, HeroesII and heavy benchmarks were not run. No boxes, downloads, access retries, agents, successors or deployment were used. Box ID/state/expiry and lifecycle: N/A; ownership: none. Free disk remained above 2 GB. Next step is coordinator review of these bounded findings; this single assignment is complete.