OpenGL stack, command and display-list crash sweep — 2026-11-01
Run: 20261101-gl-stack-command-crash-sweep (local, ignored evidence in
scratch/runs/20261101-gl-stack-command-crash-sweep/). All four requested
fixtures passed sequentially. No unexpected native crash, unhandled Wasm trap,
signal or timeout occurred. Two deliberate assert.throws(WebAssembly.RuntimeError)
guards in the display-list fixture passed. These fixtures establish neither
overall crash freedom nor gameplay compatibility.
Revision, environment and build evidence
- Requested baseline:
c08ac560f008bf74d0f40ee417f7e9bd11156b3a, branchmain. - Actual HEAD for every test:
ae6767ceb68bf39b8aae4c1c35071ef74ce94a47. The intervening commit changes onlyTODOS.mdandops/STATUS.md, allowed by the assignment's documentation-only advancement provision. - Linux x86-64, Node
24.18.1, V813.6.233.17-node.50. Local headlessrender-helperuses the repository's JS raster canvas; no browser/WebGL driver, game executable or remote box was exercised. - All 4,462 inventoried inputs match the accepted preceding run's
inputs-after.json, match baseline tracked bytes, and remain unchanged before/after these tests. The 67 distinct loaded CommonJS modules were hashed at exit and checked again afterwards; all 45 tracked loaded modules match baseline. The other 22 are installed dependencies, with exit-time byte identities recorded, not Git provenance. The sealed compiler graph's sources are covered by the input inventory. - Reused accepted
timeout 120s flock /tmp/wine-build.lock bash tools/build.shevidence from20261101-d3d-typed-stateblock-injection/build.logandoutcomes.json: exit 0, 46.479 seconds. No full build or accepted D3D test was rerun. Current primary artifact is 2,025,029 bytes, SHA-2562be3cebaa1c08a8e12075de000bfb83fb8692841c1452d9e317203f5c2fce8cf.
Commands and outcomes
Each command below had the exact prefix
SWEEP_CAPTURE_PREFIX="$RUN/<test-stem>" timeout 120s node --require "$RUN/capture-module.cjs",
where RUN=/home/user/wine-assembly/scratch/runs/20261101-gl-stack-command-crash-sweep.
The scratch-only capture hook records original Wasm bytes before instantiation
and loaded CommonJS file hashes on exit. It does not replace the compiler,
assertions, fixture or production source. Full argument arrays, start times
and logs are in outcomes.json and the four <test-stem>.log files.
| Test argument, in execution order | Exit | Seconds | Coverage kind |
|---|---|---|---|
test/test-gl-matrix-stacks.js |
0 | 8.446 | Compiled Wasm state exports versus JS matrix oracle and semantic checks |
test/test-gl-attrib-stack.js |
0 | 5.505 | Compiled Wasm observer/state exports |
test/test-opengl-command-stream.js |
0 | 0.123 | JS reference encoder, real JS replay/RPC plumbing, mocked endpoints |
test/test-opengl-display-lists.js |
0 | 6.118 | Compiled Wasm lists/encoder/software raster, plus JS replay doubles |
No assertion, production, test, shared helper or compiler file was changed. No test was skipped or retried.
Meaningful invariants reached
Matrix state: all four stacks start one entry deep; the scripted matrix
operations preserve selected stack and depth and compare state after each
step. Shipping JS identity/multiply/frustum/ortho provide direct oracles;
translation/scale/rotation use transcribed references plus independent vector
checks. Non-axis rotation uses the fixture's documented tolerance. Projection
underflow keeps depth zero (one entry); 31 pushes fill 32 entries; the next
push reports 0x0503, preserves depth and top. This boundary exercise is on
the projection stack, not an exhaustive overflow sweep of every stack.
Texture-unit matrices remain independent; invalid matrix mode leaves selection
unchanged. Lighting/material/fog defaults, clamps and selected light bounds
also pass, including modelview light-position transformation and matrix state
still working after lighting writes.
Attribute state: mutation between push/pop proves restoration of material,
shininess, ambient and active texture state; two distinct nested frames restore
in the correct order. Empty pop reports 0x0504, changes neither state nor
depth, and leaves the mirror trusted. Sixteen pushes succeed; the seventeenth
reports 0x0503, leaves depth 16, and latches the glPushAttrib untrusted
opcode. The fixture deliberately asserts the existing compatibility gap:
individual light parameters and fog are not restored. Passing is not proof of
complete OpenGL attribute-mask semantics or post-overflow recovery.
Command stream: scalar arguments and small pointed matrices survive caller
scratch reuse; opcode/argument widths, sparse-pointer vectors, immediate-mode
packing and context ownership pass. A 256-byte reference-encoder capacity
forces submission before overflow without presenting; only explicit present
publishes a frame. Texture upload borrows live guest bytes synchronously after
pending work. Query barriers submit before return. Main/worker RPC doubles
preserve order, result/output visibility and acknowledgement timing, with
optional supported/disabled/unsupported pool metadata. This test imports
test/helpers/gl-reference-encoder.js, not a compiled native producer. Its
WebAssembly.Memory objects are real shared memories, but no Wasm module is
compiled or instantiated: native encoder capacity, actual worker scheduling
and browser GPU consumption remain unvalidated by these checks.
Display lists: compiled Wasm distinguishes reserved from defined names;
reservations are contiguous/disjoint; scalar payloads survive stack reuse;
compile-only defers execution and compile-and-execute runs once immediately.
Nested playback resolves replacement definitions and ignores deleted names;
deleted names can be reserved again. Empty and maximum-GLuint names work,
self-recursion returns (source bounds it at 64), and contexts have independent
namespaces. Zero/negative/oversized reservations, invalid modes, nested
compilation and EndList during an open primitive produce their expected
errors. Reset removes both existing namespace and in-progress capture.
Unsupported pointer capture (glLoadMatrixf) deliberately traps before state
mutation; unrelated state change (glBindTexture) inside Begin/End also
deliberately traps. Both are caught by existing assertions, not new failures.
Material payload ownership, per-vertex emission and strip winding pass through the real WAT encoder with captured host batches. Invalid material inputs do not corrupt primitive/state; the injected test export verifies production stdcall cleanup and guards the following stack word. The software rasterizer actually draws a red point with the earlier material and produces no green pixels from a later material change. JS doubles additionally verify scoped lighting restoration after a draw exception. These are narrow lifetime/state checks, not allocator leak measurement, context-destruction coverage or a game.
Exact modules and reproduction evidence
The matrix and attribute runs independently instantiate identical 2,024,992-byte
modules, SHA-256
ac9beb2ae3f746c621d0f258d4d3a2d1910207b0006e13ed96856bd0ac4975d9.
All nine non-custom sections exactly match the primary artifact; the build
adds only its 37-byte encoded wine-region-layout custom section. Thus these
are actual compiled canonical-source Wasm runs, with the stamp distinction
recorded rather than claiming byte identity with the shipped artifact.
Display lists instantiate a 2,025,078-byte module, SHA-256
344f3c8096ac13e4f7d85e9cf4f3b2d8531696b66dc6fc1a0dfe8a81d5ff0a1d.
Its only fixture source addition is the original material_api_test export
appended to 13-exports.wat; GL_MATERIAL_MUTATION and render-artifact override
environment variables were absent. The original extra WAT, full transformed
exports source, their hashes, and section comparison are retained in
display-lists-extra.wat, display-lists-13-exports.wat and
module-identity.json. Three exact original .wasm captures remain ignored;
the command-stream test produces none.
baseline-inputs.json, inputs-before.json, inputs-after.json,
environment.json, loaded-verification.json, verification.json and each
test's module/loaded-module records supply full source/module identities.
Reproduction uses the four unchanged entrypoints and the commands above on
matching input bytes. No crash reproduction was needed.
Remaining evidence gap is native producer overflow/order under actual consumer load and real application gameplay; neither is inferred from this sweep. REP/held tasks, Heroes II, browsers, benchmarks, downloads, access retries and deployment were untouched. Free disk stayed above 2 GB (2,446,188,544 bytes at post-test verification). Box ID/state/expiry and lifecycle: N/A; no box assigned or accessed, ownership none. Next: coordinator reviews this single completion; this worker stops without launching a successor.