Independent allocator-reuse review and POSITIONT COM qualification

ACCEPT_FIXTURE; full retained-module COM qualification PASS. Review was completed and recorded before native execution. Four fresh sequential Node processes produced baseline exit 0 and all three intended control exits 1. No compilation, production/fixture repair, donor mutation or retries occurred.

Independent review

Reviewed candidate c9a59870ff46416653d0782ebfc296df1bb0ae59 against 74377b3382e03a7ead1a3e33981321b768d6c3ca. The exact fixture diff has SHA256 c1a61f4c72eb1e2d3ed3674694ae75641ed97a6ba7449272088138b02334f3d6. Fixture blob 4ea43dde7 (full blob in source manifest), SHA256 6fea929b37398641c4fdc329e42d4e46ee739ddd65f28e4ad5771b0592fff93d, remains mode 100755. All 23 predecessor seals were independently checked against both retained files and committed bytes. The predecessor report, original, focused and no-op raw logs/receipts, causal summary and closure-after were read. Those three runs were not repeated.

The source ordering establishes why the correction is valid:

• lib/d3d-shader-ir.js:75 copies native bytes with .slice() before returning readFixedBundle; fixedPositionT finally frees the bundle and descriptor. src/09aj-d3d-fixed.wat:136 frees the four subordinate allocations and bundle; src/09ag-d3d-shader-vm.wat:92 converts linear pointers with real w2g.
• lib/d3d9-host.js:1468 puts that independently owned IR into the actual production payload before _issue. The fixture copies this payload for its hash oracle, then poisons real allocator returns before delegating _issue. Thus the hash alone observes a task copy, but the subsequent authentic draw independently consumes the production payload and checks complete BGRA/depth. It does not rely on the task copy to establish real-draw non-aliasing.
• reuseProof is installed immediately before the synchronous recovery draw. The DRAW wrapper saves it and clears it before invocation. There is one such draw before the null assertion; source control flow proves exactly one call. An exception becomes DrawPrimitiveUP failure and cannot reach baseline PASS.
• src/10-helpers.wat:2868 uses aligned extents, exact/next-larger bins and first-fit free blocks, splitting from the tail. Free enqueues blocks without coalescing. lib/d3d9-software-backend.js:495 coalesces during later prepare; src/09ai-d3d-render-lifetime.wat:159 merges only free adjacent blocks. Original evidence shows frees 210→215→216, then 17 merges and backend reuse: historical bundle start 0x1010564 now has 12 B payload rather than 36 B. Requiring its old size there after backend work was an invalid fixture demand.
• Before prepare, descending target sizes prevent this search from carving a smaller pending target to satisfy a larger request. Equal-size targets are recorded in a set regardless of allocation order. Poison covers only successful fresh guest_alloc(size) payloads; these are held in live until queue proof completes, so allocator-valid live production owners cannot be overwritten. No arbitrary historical address is directly poisoned. This conclusion is for the pinned allocator and this fixture, not a general heap-layout guarantee.

The exact diff changes only callback placement and descending target ordering. 64 attempts per target, 1 MiB cumulative requests, real allocator/free, six exact addresses, poison 0xa5, IR hashes, full pixels/depth, queue and final-release oracles remain. Production, all 286 pins and all 290 closure file modes are unchanged relative to the candidate. Original fixture acceptance 7aef3ff and aea185, guard/boundary, 52 gates, decoder, arithmetic and build evidence are reused; none were replayed.

No-op mismatch disposition: the mutation skips the real five-block recovery bundle free while still freeing the descriptor. Its trace stays at free count 210 for the bundle, leaving the largest target IR live. The real allocator returns other blocks for 63 requests of 16420 B (1034460 B); the next request would exceed the unchanged 1 MiB bound. Native exit 1 at that proof assertion is causal rejection of missing free, not generic crash or external exhaustion. The runner also exits 1 because it expected the narrower exact-address message. That runner mismatch is retained and is not a runner PASS. It limits the prior runner receipt but does not invalidate the bounded proof: skipping free cannot satisfy the unchanged exact-address obligations within its byte budget. No prior full qualification was accepted.

Exact retained inputs and commands

Read-only module donor: /home/user/wine-assembly-render-positiont-tiny-rhw-native-qualification-oct23/build/positiont-com-tiny-rhw.wasm, 1883361 B, SHA256 5d057b5424b710fb28f1c617786ba71d7c8c116fb74e51eac611ec9a79868ef1. Adjacent positiont-com-tiny-rhw.provenance.json, 1002 B, SHA256 e64c83f925f52649f46e1a89bb952d7dde06d627f7a799f0956a535c1bca8cc3. Pins digest d39a81a012129bd7c81f4f697d0a8317b3869a8195fc6fd32c1ff13fc3431262; extraWat digest 4bd65ac87452abb4ec5f728f88b399439c9fa4fec91a53b3e6cb96468f58d068. Tiny-RHW source remains reviewed production 301b568d, review 345ea682, single source blob 538290c4241f9112ff5e16f5cd71735ab38edea9. The 290-file manifest preserves pinned source identity, physical bytes and modes.

The module loader requires explicit module/provenance, validates both and all source pins. Its separate buildDiagnostic is unreachable here. render-helper selects prebuiltModule before its lazy compiler branch and creates exactly one shared 8192-page Memory and one Instance; no Workers. The accepted guard /home/user/wine-assembly-shadow-man-native-bound-canonical-oct23/docs/re-notes/shadow-man-native-bound-canonical-20261023/native-module-bound.cjs SHA256 47ce2e2b835be33eab6b2b24f027ecaab8ceb61c559890e51548c78199beb197 synchronously checks intrinsic BufferSource lengths ≤8388608 before native Module construction. It is loaded by --require, unmodified.

Each full argv is recorded in its JSON receipt. Installed Node /home/user/.nvm/versions/node/v24.18.1/bin/node --max-old-space-size=512 --require GUARD test/test-d3d9-positiont-com-qualification.js --qualify --allow-512m --module=ABSOLUTE_RETAINED_MODULE --provenance=ABSOLUTE_RETAINED_PROVENANCE runs first without a control, then with --negative-control=disable-shared-route, --negative-control=alias-queued-bytes, --negative-control=corrupt-queued-ir. NODE_PATH=/home/user/wine-dashboard-deps-20261014/node_modules uses the independently measured pinned pngjs in place. Compiler/cache/debug overrides are cleared. No instrumentation changes the qualification fixture.

Baseline reaches tiny-RHW fallback complete pixels/depth, real descriptor allocation rejection/recovery, all six reused addresses and 34164 requested bytes, poison/hash, queue mutation survival, both invalid ranges, real resource and device final Release, destroyed backend, zero backend bytes and advancing production frees. Final counters are 351 allocations/281 frees; they are not a claim that every allocation in the complete harness is freed. Controls reject at intended assertions rather than crashes. Raw logs and assertions are retained.

Admission and handback

The consolidated measured admission was 46918232/67108864 B, including 27367589 B materialized closure, private Git/index/new objects, directories, scripts/evidence/temp/captures, external launcher events/stderr, separate Codex session JSONL, 6 MiB future growth and 8 MiB reserve inside the cap. Host after projected growth was 5777352539 B, above 2982150144 B. Standing donor, module, executable, pngjs and direct object-store sizes/hashes are separately recorded. No links conceal copies; no fetch or archived recovery.

Fresh MemAvailable before the four commands was 3050934272, 3036528640, 2946469888 and 3019022336 B, each above 1610612736. Each command had CPU/wall 150 s, 512 MiB old space (not an RSS limit), regular-file limit 24 MiB, aggregate capture limit 512 KiB, fail-closed 125, TERM+5 s KILL and Linux subreaper/full reap. Total native capture 9555 B; peak child RSS 97320960 B. All four receipts show no guard failure, groupReaped true and no adopted descendants. The heavy arena is released. Publication verification and final accounting are retained externally in the task publication receipt.

Only this report and adjacent evidence are new relative to the candidate tree. No shared integration/main/merge/deploy, compiler/build, fixture/production repair, donor mutation, Boat/browser/GPU/guest, worker/refill or other-process intervention. Shadow Man 01a16770 Python-only repair, RichEdit, Claude B&W2 and laptop ownership are preserved. Box ID/state/expiry N/A; lifecycle coordinator. No performance, WebGL or game-acceptance claim.

Git does not back up the ignored module. Recovery requires the exact retained read-only artifact reference above, hashes/provenance and source manifest in this evidence directory. A future authorized reproduction must materialize those bytes/modes in a fresh isolated root, take fresh admission and use explicit retained module arguments; no rebuild is authorized by this recipe. This task's four native processes are consumed. Stop; coordinator assigns any next task.