Compiled-Wasm OpenGL encoder crash sweep — 2026-11-01

Run 20261101-gl-native-encoder-crash-sweep: all three requested fixtures passed sequentially with their original assertions. The first ran every section, without a --section filter. No unexpected native crash, unhandled Wasm trap, signal, assertion failure or timeout occurred. Six intentional native trap assertions passed; they are rejection coverage, not crashes. Evidence remains ignored in scratch/runs/20261101-gl-native-encoder-crash-sweep/.

Revision and build identity

Requested baseline is c4472ee3bda7dc4b87549c4f594662206a7e15d7 on shared main. Actual HEAD for all three commands was 62211324e4fa6e0465199513dec92abb6f3cf61b; the intervening change touches only TODOS.md and ops/STATUS.md, within the permitted documentation-only advance. All 4,462 inventoried inputs match baseline tracked bytes, match the accepted build run's inputs-after.json, and are unchanged before/after this sweep. All 67 distinct loaded CommonJS files were hashed at exit and checked afterwards; 45 tracked loaded files match baseline. The remaining 22 installed dependencies have recorded byte identities, not Git provenance. The inventory includes the sealed compiler graph used by render-artifact.

Reused accepted build evidence from 20261101-d3d-typed-stateblock-injection/build.log and outcomes.json: timeout 120s flock /tmp/wine-build.lock bash tools/build.sh, exit 0 in 46.479 seconds. No build or accepted test was rerun. The current primary artifact remains 2,025,029 bytes, SHA-256 2be3cebaa1c08a8e12075de000bfb83fb8692841c1452d9e317203f5c2fce8cf. Environment: local Linux x86-64, Node 24.18.1, V8 13.6.233.17-node.50; no WINE/GL/NODE_OPTIONS override was set. Memories are real fixed 512 MiB shared Wasm memories. No browser or GPU driver was exercised.

Commands and outcomes

Each test argument below followed the exact prefix SWEEP_CAPTURE_PREFIX="$RUN/<test-stem>" timeout 120s node --require "$RUN/capture-module.cjs", with RUN=/home/user/wine-assembly/scratch/runs/20261101-gl-native-encoder-crash-sweep. The scratch-only hook captures original module bytes before instantiation and loaded CommonJS hashes on exit; it changes no compiler, source, assertion or fixture. Full argument arrays, timing records and logs are retained in outcomes.json and <test-stem>.log. The encoder timing record labels its post-command timestamp completedUTC; the other two record startUTC.

Test argument, in execution order Exit Seconds Native coverage
test/test-opengl-wat-encoder.js 0 3.245 All encoder parity, pressure, boundary, error, isolation and pool sections
test/test-opengl-sparse-page-arrays.js 0 3.091 Sparse index/vertex reads, upload snapshots and software texture state
test/test-opengl-client-attrib.js 0 3.729 Client stack state, packed draws, context/list behavior and host replay

No retries, source repairs, assertion changes or skipped sections were needed.

Meaningful coverage and its limits

Native producer and independent oracle. The first fixture calls gl_wat_encoder_call and gl_wat_stream_flush on actual compiled Wasm. Its separate JS reference encoder provides expected bytes. Non-pressure comparisons require identical submissions and bytes; pressure comparisons normalize only packet-relative payload offsets and require every complete record in order. The mocked gpu_gl_call captures bounded native batches, checks record walks, and supplies deterministic context/query results and output writes. The initial shared-offset range/alignment/malformed-record checks exercise JS replay, not a native trap or actual worker scheduler.

All ten primitive modes run smooth and flat, including incomplete counts 0–5. State checks cover current attributes, vector inputs, server attribute push/pop, both texture units and projective r/q. Strided client arrays cover eight scalar types, indexed draws and offset/empty DrawArrays. Snapshot and barrier checks include copied matrices/vectors, borrowed uploads, query output visibility, 22 queued small texture uploads submitted together at finish, mutation after capture, the 65,536-byte copy threshold, and synchronous large/padded/unsupported upload cases.

Capacity and pressure. The unchanged production capacity is 2 MiB. Three cases exceed it over the call sequence: generic records, packed triangle records, and a large packed primitive following an occupied stream tail. The first two explicitly require native submissions to outnumber the 2 MiB oracle's submissions and the first batch to be within 300 bytes of the 512 KiB production watermark. All three require complete ordered-record parity across batch boundaries. This proves native producer flushing in the fixture; it does not measure real GPU/worker consumer backpressure, scheduling or throughput, nor exhaustively reject every possible oversized record.

Hardening and memory boundaries. Context transitions include failed make-current and failed deletion, 24 context keys including zero, reverse revisits and deletion. Twenty-four nested server attribute frames, typed integer extrema, NaN, negative zero, tiny doubles, unknown vertex types, disabled pointers and configured pointer zero pass parity. Vector inputs at the actual memory end exercise exactly three color floats and two vertex floats; a direct frame containing return address plus exactly two arguments ends at memory end. The fixture's original translator sentinels place these vectors there; this is a targeted overread check, not general pointer safety.

Six invalid inputs deliberately require a JS RangeError and a caught native WebAssembly.RuntimeError: nested Begin, generic command within Begin, ShadeModel within Begin, unknown opcode, 64-bit texture-byte overflow and DeleteTextures count-shift overflow. Reset separates these cases. Reset/default state and two Wasm instances interleaved over one shared memory also pass; the instances run sequentially, not on concurrently scheduled worker threads.

Stream pool. Original allocation hooks exercise successful triple-buffer allocation, failure of either extra buffer, and default non-pooled behavior. Assertions cover disjoint buffers, two published buffers retaining bytes while the third is written, rotation/host return values, empty flush, invalid and late configuration, partial-allocation rollback, and no allocation retries. Reset preserves the current buffer. A deliberately thrown JS host-publication error is caught; pending bytes and buffer selection survive, and retry succeeds. That host exception is neither a Wasm trap nor a native crash. Retained buffers are checked with a mocked consumer; real lease acknowledgements are absent.

Sparse client inputs. The second fixture maps five sparse pages in an interleaved order and asserts backing pages are nonadjacent; unrelated backing contains 0xEE. Nine uint16 indices and four float vertices yield the same expected packed positions inside a page and across two guest-page boundaries, including a vertex whose components span pages. A 1,024-byte queued texture snapshot crosses nonadjacent backing pages and survives guest mutation. A four-float border color spanning a page boundary reaches actual native software texture state and copied replay payload, also surviving mutation. These are selected mapped-page crossings, not unmapped-page fault coverage, all scalar widths or arbitrary byte-unaligned scalar straddles.

Client attributes. The third fixture's real native encoder restores all supported vertex/normal/color and two-unit texture descriptors, strides, enables and pointer validity, producing an identical 3×19-float packed draw. Selective masks distinguish vertex state, client unpack modes and server current color. Restored unpack alignment reaches the host frontend and the 14-byte RGB upload. Zero/unsupported masks, all bits, depth query, 16-frame limit, overflow/underflow GL errors, Begin/End misuse, context isolation, deletion cleanup and immediate client operations within compiled lists pass. These GL error values are asserted returns, not native traps. Actual JS OpenGLHostBridge/FixedFunctionGL replay is used with mocked backend creation and upload capture; no real GPU draw or application gameplay occurs.

Exact compiled modules and reproduction

Original captured module identities:

Fixture Bytes SHA-256
WAT encoder 2,025,196 83b7954c2e55f32c8d716f478e294232db033b42c64ca015e0b6e941aa41afde
Sparse page arrays 2,025,050 c1f9f9c16d8ccac95f598f3936c45471767b1fc5450bf2acab92701cc7ec99e1
Client attributes 2,024,992 ac9beb2ae3f746c621d0f258d4d3a2d1910207b0006e13ed96856bd0ac4975d9

Client attributes has all nine non-custom sections identical to the primary artifact; only the build's 37-byte encoded wine-region-layout custom section is absent. The other modules intentionally differ through the unchanged fixtures' original compileSrcWasm transforms. Encoder replaces affine allocation/free call sites in 09a8c-gl-encoder.wat with counting/fault hooks and redirects 09a8e-gl-state.wat translation through the two memory-end sentinels (ordinary addresses still use production g2w). Sparse arrays appends only test_sp_map and test_border_ptr exports to the state fragment. These are compiled native producer tests with explicit instrumentation, not byte-identical shipping-module tests.

module-identity.json contains section hashes and full module identities; three exact original .wasm captures are retained. verification.json records full fixture source hashes; <test-stem>.original-source.txt preserves their exact code, including transforms. inputs-before.json, inputs-after.json, loaded-verification.json and environment.json record source/dependency identity. Reproduction uses these same unchanged entrypoints and command arguments on matching source bytes. No unexpected crash reproduction was needed.

Remaining gaps are actual consumer-load/concurrency behavior, browser GPU execution, arbitrary invalid memory inputs and application gameplay. This sweep establishes no full-27 result or general crash-free claim. REP tests, diagnostics, automated-review held work and Heroes II were untouched; no browser, heavy benchmark, box, download, access retry, worker or successor was used. Post-test free disk was 2,434,781,184 bytes, above 2 GB. Box ID/state/expiry and lifecycle are N/A; no box ownership to hand back. Next: coordinator reviews this single completion; this worker stops.