Graphics and fault native sweep, 2026-11-01

Five assigned harnesses passed; the threaded REP matrix failed its first case. All six ran sequentially on exact main c2835c991da9624a112dacf4e095ef26d5dc7308, tree 910dbf41b9f35bdeb037f58d3ad214f8017342c0. Evidence run: scratch/runs/20261101-graphics-fault-sweep/. No production or test source was edited. The previous seven-test sweep was not repeated.

Test in test/ Exit / seconds Result
test-d3d-device-creation-oom.js 0 / 5.172 PASS: D3D2/3/7, 24 allocation failures and 24 valid retries; no partial publication, heap/object leaks, retained creator or slot consumption; ABI checks.
test-d3d9-shader-retention-oom.js 0 / 4.520 PASS: three final-allocation failures clear output, preserve refs and release both temporaries without IR leak/double free; valid retry/readback/release.
test-seh-unhandled-terminates.js 0 / 0.165 PASS: unhandled guest stops once at EIP 0/yield 2; a continuing raise storm yields to the host.
test-unhandled-exception-filter.js 0 / 4.645 PASS: setter chaining/clear, vectored registration/removal and stack cleanup, null pointer/record survival. Printed exception reports are asserted input paths.
test-fault-reserved-page-cross.js 0 / 4.331 PASS: dword read/store retranslate when a synchronous filter decommits the first page.
test-rep-movs-fault-restart.js --matrix 1 / 4.825 FAIL: first of 80 cases, w1/df0/source/k0/cold/aligned/plain/mode3/threaded; handler called 7 times, expected 1. Remaining 79 cases were not reached.

The REP failure also reproduced without the evidence preload using node test/test-rep-movs-fault-restart.js --width=1 --operand=source --completed=0 (exit 1, 4.768 seconds). Both runs compiled module SHA-256 364f592ab6d8e9fd2f8816046427616e09064cc8ccc6a3e6fbf4b8c1d2656126. The last handler observation reports EIP/ExceptionAddress 0x40e000 (4251648), whereas the REP prefix is at 0x40e00c (4251660). The pre-REP marker ends at 2 instead of 1; the post marker is 1. The repaired page is mapped, guest EIP ends at 0, and there are no unhandled guest exits. This is a reproduced restart correctness failure, not a native trap or hang. The log explicitly describes its saved context as the last handler entry; it does not establish the contents observed at the first of seven entries.

Source inspection supports a production limitation rather than a missing dependency or stale build: src/05b-string-ops.wat's $rep_movs_mem element loop does not stop after $gl8 redirects execution on a mode-3 miss, and $rep_movs_do publishes register progress only after the complete copy. src/03-registers.wat's $g2w_miss explicitly documents that asynchronous mode 3 returns sentinel scratch and allows the faulting operation to complete. The observed block-entry EIP also cannot provide precise REP restart. These facts explain the repeated dispatch and block rewind; they do not establish that every matrix case fails. No actionable native crash was reproduced, so the conditional crash-repair path was not invoked. Precise REP restart remains unrepaired and needs a separately scoped source/decoder/accessor design and regression work; no assertion was relaxed or failure reported as success.

All harness prerequisites were reviewed before execution. These are synthetic native tests with 512 MiB shared Wasm memory, canonical WATX source compilation and existing setup exports/allocation injections. The termination harness loads the existing build and notepad fixture. The filter test uses the render helper's default bundled fonts; every required font was present. The page-crossing test selects fonts: 'none'. Canvas uses the repository's pure JavaScript raster implementation. There were no dependency failures, heavy runtime skips, browsers, boats, games, downloads or benchmarks. Every process had a 120-second timeout; no process timed out. Only the default threaded REP engine was selected; optimized engines, zero-count and alternative handlers were not run.

Linux x64 Node 24.18.1, V8 13.6.233.17-node.50, default engine tiering; no WINE_* environment overrides or NODE_OPTIONS. run.py records commands, times and disk checks in outcomes.json; the evidence-only preload capture-module.cjs copies/hashes bytes then forwards unchanged arguments to native WebAssembly.instantiate. Captured module identities:

Test suffix SHA-256
d3d-device-creation-oom a49542126ec6da0a094d233a933c04450e0f2dcdbce88b85f3d84fb49de36822
d3d9-shader-retention-oom b76228f6f13cac92ae20e4b227363b45e15444d38a30352e2492d5d1756e6dab
seh-unhandled-terminates 2be3cebaa1c08a8e12075de000bfb83fb8692841c1452d9e317203f5c2fce8cf
unhandled-exception-filter 64668e162d8151b05d58802e603fb451549374d1a949e4228585c7b7e0c329a6
fault-reserved-page-cross 3083c0cf810567c1e6ab39f409602a30a05572bf173f639cf5aea702a9fbd812
rep-movs-fault-restart 364f592ab6d8e9fd2f8816046427616e09064cc8ccc6a3e6fbf4b8c1d2656126

inputs-before.json / inputs-after.json match without changes. All 369 tracked inventoried source/compiler/host/helper/harness/font files match the requested Git revision. The broad inventory also hashed foreign untracked lib/wininit-boot.js; it was untouched. The assigned REP harness is itself untracked, so its result is local harness against exact-main production, not proof that the harness belongs to that revision: 23472 bytes, SHA-256 dd1cb1ccc3e64f20b81adb1e6bf8f1895e1f46276125a76afc2b6e03047be019. Its unchanged bytes are preserved as rep-harness-original.js. The ignored notepad fixture is 53248 bytes, SHA-256 e7b01ab85a35ba6f577067d13efa522dc3d605de7712340e4acaf7065f866340, preserved as notepad-fixture.exe.

canonical-stamped-identity.json verifies the existing 2025029-byte build against fresh canonical compilation plus the build's region-layout section: byte-identical SHA-256 2be3cebaa1c08a8e12075de000bfb83fb8692841c1452d9e317203f5c2fce8cf, layout 3a03507bae8914c0, matching the generated JS mirror. This compilation check is not a new full build-gate pass. Evidence includes individual logs, captured modules, modules.jsonl, prerequisites, environment, source comparison, the reduced REP observation and reproduction scripts; all remain ignored.

Documentation-only publication uses git diff --check; no build gate or census was changed. Free disk remained above 2 GB (minimum recorded about 2.53 GB). No boxes were assigned or accessed: ID/state/expiry and lifecycle are N/A. This assignment ends with five passes and one precisely bounded failure; gameplay, full27 coverage and all-app crash freedom remain unproven.