D3D resource crash sweep, 2026-11-01

The original six requested native Node/Wasm tests ran sequentially: five passed and device enumeration failed a stale capability assertion. No native trap, process signal, timeout or production crash was observed. No source repair or test edit was made.

Baseline: 22992d60efb01985e5c55f9fdd2572e5c3732657. Actual HEAD for all six tests: f3fedefac7271983e6a02321629620efabcbd283 on shared main. The intervening changes affect only TODOS.md and ops/STATUS.md. The new runner records actual HEAD and verifies documentation-only advancement and input identities; it does not require HEAD equality.

Test (test/) Exit Seconds Observed coverage
test-d3d-root-lifetime.js 0 5.336 D3D1/2/3/7 references; six device-parent release orders, ABI and heap balance
test-d3d-render-lifetime.js 0 18.679 Direct/sparse retirement, adoption, coalescing, malformed lists, bins, passive diagnostic preserving complete memory and mutable globals
test-d3d-device-enumeration-lifetime.js 1 3.687 All 16 nested version pairs completed; failed during first D3D1 cycle, before its first callback resume
test-d3dim-execute-buffer-lifetime.js 0 3.952 Eight cached/uncached public creation, Lock, reference and final-release cycles; payload/cache balance and ABI
test-d3dim-texture-release-arena.js 0 4.222 Texture/Texture2 teardown, raw-free negative control, color-key refund, 100 balanced arena cycles
test-d3d9-depth-release.js 0 6.158 Actual x86 parked depth/backbuffer release and Reset, delayed failure/retry, immutable resize snapshot, balanced RET and shared serial exhaustion

Enumeration failed at test/test-d3d-device-enumeration-lifetime.js:77: legacy filter caps unchanged, actual 195 (0xC3), expected 255 (0xFF). Compilation, instantiation, fixture loading, nested callbacks and their allocation-balance assertion had succeeded. Static review identifies a test expectation mismatch: $fill_primcaps in src/09a8-handlers-directx.wat:3629 explicitly excludes unsupported legacy mip filters and writes 0xC3. The same test's later real-x86 assertion at line 157 already expects legacy 0xC3. This is a harness assertion failure, not evidence of a resource-lifetime production crash. The 96-cycle sweep did not complete, and all 32 later real-x86 nested enumeration cases remain unexecuted. No retry or diagnostic execution was needed for this classification.

Evidence is ignored under scratch/runs/20261101-d3d-resource-crash-sweep-run2/. It contains the new run.py, per-test merged stdout/stderr logs, commands, elapsed times, exit codes, environment, HEAD identities, static contract excerpt, outcomes and result.json. Each command used timeout 120s node --require <run>/capture-module.cjs test/<name>.js. The hook preserves the bytes and SHA-256 of each byte-based Wasm instantiation, then delegates unchanged. Ten byte captures cover six distinct test modules; the depth test also instantiates its existing compiled module for a sibling instance. module-identities.json, per-test .modules.jsonl and .loaded-modules.json retain exact test-module and loaded dependency identities.

The prerequisite inventory was reused from scratch/runs/20261101-d3d-resource-crash-sweep/ and freshly checked: all 1,167 source/compiler/host/helper/test/fixture/build files match the prior inventory and remain unchanged before/after execution. All 1,163 tracked inputs match baseline Git blobs; the other four are ignored notepad/calc fixtures and the primary/compatibility builds. All 46 loaded repository module hashes also match baseline. Full SHA-256 and sizes are in inputs-before.json, inputs-after.json and the identity summaries. The existing primary build remains 2,025,029 bytes, SHA-256 2be3cebaa1c08a8e12075de000bfb83fb8692841c1452d9e317203f5c2fce8cf. Each test uses the canonical WATX closure/compiler through the render harness with its existing test exports and fonts: 'none'; no artifact override, compiler/census bypass or new full build is claimed. Both PE fixtures exist; no missing dependency blocked execution.

The first attempt ran no tests: its obsolete exact-HEAD assertion rejected coordinator documentation commit e58f596a4b33fe72d28ed79ab3fd2a34b4558281 against requested 00072255dd3d38df6e4d87935da9b1d9a7f17dd7. Its original prerequisite/failure evidence remains intact in the prior run directory. The present results replace that blocker under the explicit identical-input, documentation-advancement allowance.

Environment: Linux x64, Node 24.18.1, V8 13.6.233.17-node.50; no WINE_* or NODE_OPTIONS overrides. Free disk stayed above 2 GB (final 2.506 GB). Only this findings document is committed; documentation diffcheck is the publication check. No production/test/compiler/census, TODO/STATUS or foreign WIP files were edited. No REP tests or diagnostics, held-topic work, browser, heavy runtime, benchmark, download, access retry, boat or worker was used. Box ID/state/expiry and lifecycle are N/A; no box ownership to hand back.

Next: separately scoped reconciliation of the enumeration test's stale capability expectation, followed by its remaining coverage. No gameplay, full27 or all-app crash-free conclusion is supported. This worker stops after publishing this single task.

Enumeration assertion reconciliation

Follow-up run: 20261101-d3d-enumeration-assertion. The one-line legacy callback assertion now requires exactly 0xC3 instead of 0xFF, with a message identifying the excluded legacy mip filters. No lifetime, ABI, ordering, cancellation, canary or allocation assertion was removed or weakened. All 16 nested version pairs, 96 balanced cycles and 32 real-x86 nested COM/continuation cases now complete (exit 0, 5.246 seconds).

This correction follows the existing contract rather than changing production: commit 7fec2dd33634f7ce76950cdecec8870ea29508d3 deliberately changed $fill_primcaps from 0xF3 to 0xC3, excluding legacy mip bits 0x04..0x20. That commit updated this harness's D3D7 and real-x86 expectations but left line 77 at 0xFF. Current source, the later real-x86 assertion and both extracted filter regressions require legacy 0xC3. The unchanged harness first reproduced the exact line-77 failure (195 !== 255, exit 1); the corrected harness then passed in full.

Validation has a separate, pre-existing blocker: the unmodified test-legacy-texture-filter-caps.js fails at line 121 and test-legacy-final-mip-bits.js at line 126, both reporting texture ops unchanged, actual 0xF, expected 0x3FF. Their legacy filter and preceding descriptor checks pass, but their later checks and mutation negatives do not complete. The final-mip test also prints its 36 backend state cases before failing. Neither regression is claimed as passing. Production explicitly advertises 0xF at descriptor offset 180; history f42083c61a18d56fd739d3cd702c4d67d36784b7 documents the accepted DISABLE/SELECTARG1/SELECTARG2/MODULATE mask. The unchanged newer test-legacy-texture-operation-caps.js --negative-controls passes its 9 pairs, 144 argument combinations and 15 rejecting controls. No production contract bug was established. Repairing the older regressions is outside this task's file ownership; their stale operation-mask checks remain open.

Mandatory flock /tmp/wine-build.lock bash tools/build.sh passed in 48.565 seconds, wrapped in timeout 120s. All native test commands also used 120-second timeouts. The five previously accepted resource tests were not rerun. No REP tests or diagnostics or other held-topic work ran.

Baseline was 4e67faf1ea2e6d435c7fae93c51eba549ae0bf79; initial HEAD matched it, and subsequent tests ran at 62693ec240f9b8842e2be796da0b09b3f34de5b5, whose only intervening changes were TODOS.md and ops/STATUS.md. The 1,169-file input inventory matches baseline/prior evidence before correction; afterward only the owned enumeration test changes, including byte-identical primary/compatibility builds. Exact per-file SHA-256, loaded module identities, commands, exits, logs, source/history excerpts and the minimal patch are retained in the ignored run directory. Before/after enumeration captures are identical: 2,025,153 bytes, SHA-256 001c726c332311bf570ebf9fff67e04608c407d7ff3f80c5c3f9aba5e0611617. The rebuilt primary module remains 2,025,029 bytes, SHA-256 2be3cebaa1c08a8e12075de000bfb83fb8692841c1452d9e317203f5c2fce8cf.

Linux x64 / Node 24.18.1; no WINE_* or NODE_OPTIONS overrides. Disk remained above 2 GB. No source/compiler/census, foreign WIP or TODO/STATUS edits, browser/heavy runtime, download, access retry, deployment, agents or boxes. Box ID/state/expiry and lifecycle are N/A. Coverage is limited to the named harnesses, with the two caps-regression failures explicitly unresolved. Next: separately scoped correction of those regressions' stale operation mask assertions; no broader compatibility claim follows from this run.

Filter and final-mip assertion reconciliation

Run 20261101-legacy-filter-assertions resolves the two older regression failures above. Each test changes only its D3D7 descriptor offset-180 expectation and message: 0x3FF becomes exactly 0xF, identifying DISABLE/SELECTARG1/SELECTARG2/MODULATE. All filter, backend, guard, ABI and mutation assertions remain intact. No production change was needed.

History f42083c61a18d56fd739d3cd702c4d67d36784b7 deliberately added DISABLE to the previous 0xE mask. The current 236-byte D3D7 descriptor stores 0xF at offset 180, followed by the two 16-bit stage/texture limits at 184/186; this field is separate from the line/triangle filter caps at 40/96. The production combiner terminates stage zero on DISABLE and handles SELECTARG1/SELECTARG2/MODULATE; the newer regression checks those operations, unsupported-operation fallback, argument handling and GPU lowering. This contract and the prior recorded failures justify the correction.

All three native tests ran sequentially with timeout 120s node --require <run>/capture-module.cjs test/<name>.js; only the operation test takes --negative-controls. The older tests execute their negative controls unconditionally (their optional positional argument is a DirectX source path).

Test Result Completed coverage
test-legacy-texture-filter-caps.js PASS 90 descriptor fields, 36 software/GPU filter cases, point/bilinear footprints, all 6 mutation negatives
test-legacy-final-mip-bits.js PASS 90 descriptor fields, 36 backend cases, real WAT pixel fetch/decode on two bound images, all 10 mutation negatives
test-legacy-texture-operation-caps.js --negative-controls PASS 9 operation pairs, 6 unsupported-operation fallbacks, 144 argument combinations, all 15 rejecting controls

Mandatory timeout 120s flock /tmp/wine-build.lock bash tools/build.sh passed in 44.589 seconds; focused tests took 0.278, 0.292 and 0.639 seconds, respectively. Own-path diffcheck passed. The primary build remains byte identical: 2,025,029 bytes, SHA-256 2be3cebaa1c08a8e12075de000bfb83fb8692841c1452d9e317203f5c2fce8cf.

Baseline 168c9a6369253f7627bb5d6f105888020ed0fd6a and tested HEAD 91c02f1c36059b06a866879f6a3b81a12254fd33 differ only in TODOS.md and ops/STATUS.md. Before correction, all 1,169 inventory files match the prior run's final inventory and relevant tracked inputs match baseline. Ignored evidence retains exact commands, outcomes, environment, before/after input hashes, loaded module identities, captured Wasm bytes and SHA-256, source and history excerpts, and the patch. The capture hook delegates module creation unchanged and records mutation modules as well as the passing fixtures. After validation, only the two owned tests differ across the input inventory; all 10 loaded repository modules match baseline except those two corrections.

Coverage remains extracted production source with fixture memory and GPU descriptor/lowering checks, without a graphics context or full guest runtime. Enumeration wiring is checked statically here; accepted lifetime/enumeration tests were not rerun. No additional production issue was observed. No REP tests or diagnostics, held-topic work, foreign WIP edits, compiler/census changes, browser/heavy runtime, boats, downloads, access retries, agents, successors or deployment. Disk stayed above 2 GB. Box ID/state/expiry and lifecycle are N/A, with no runtime ownership to hand back.