D3DIM texture-wrap recording fixture repair — November 7, 2026

Run: scratch/runs/20261107-graphics-wrap-fixture/. This repairs only test/test-d3dim-texture-wrap.js. Production rendering is unchanged.

The unchanged test was executed first (baseline-test.log, exit 1). Its software16 assertions completed, then its recording device failed with Cannot read properties of undefined (reading 'gl') in _submitImmediate. The fake declared a 16-bit target but implemented only device.draw; production correctly calls _storeNative16 after each primitive.

The fixture now explicitly observes _storeNative16 on its own executor instance. It asserts the target identity and absence of a real GPU, and requires alternating draw/storage events for every triangle. Each submission must contain one triangle, three 32-byte vertices. The UV oracle reads all submissions, respecting each typed array's byte offset and length. Reading draws[0].vertices.buffer previously depended on the backing buffer also containing other submissions, rather than checking their submitted views.

All existing software pixel checks, describe field 33, WRAPU/WRAPV/WRAP0 expectations and per-triangle UV expectations are unchanged. No production fallback or conversion bypass was added. This recording fake validates UVs and submission boundaries; it cannot validate native16 pixel conversion.

Validation and limits

The actual native16 regression exercises ordinary public COM draws and native Lock readback in both shipped modules, direct WebGL and the owning queued Worker. All 16 positive configurations / 96 dependent pixel pairs passed with zero native16 tolerance. All 56 negative controls rejected their alternative pixel oracle: stale destination, conversion only at final readback, conversion only per public draw, and stale alpha where applicable. Same-state merged draws and overlapping primitives distinguish per-primitive storage from deferred storage; depth/scissor cases also remain covered. The negative errors in native16.log are expected assertions verified by the unchanged regression, not positive-run failures.

Local Chrome was 151.0.7922.108, launched headless with SwiftShader flags. Actual contexts reported WebGL 2.0 and ANGLE (Intel, Mesa Intel(R) UHD Graphics 620 (KBL GT2), OpenGL 4.6 (Core Profile) Mesa 23.2.1); launch flags alone do not establish backend identity. All positive cases recorded zero errors and zero fallbacks. native16/result.json records renderer identity, native observations, successful public calls, merging/upload/readback counts, both loaded module hashes and normal browser exit (code 0, no signal). These are actual WebGL API/pixel checks on the recorded backend, not physical-GPU, native Windows, original-game, performance, or full software/WebGL wrap parity claims. No UV discrepancy was found in the existing fixture. No original title or remote box was used.

Exact source and modules

Unchanged reproduction base: baseline-sha.txt. Tested source base: 492cf5f375b4393918023acfcedb1f459444e6e0 (tested-source.json). tested-source.patch records relevant tracked uncommitted changes with git diff --binary; fixture.patch isolates this repair. The source manifest pins tracked source/build tools and relevant test helpers, including a broader inventory than the executed dependency closure. Unrelated shared-tree work is not part of this repair. focused-harness-identity.json seals the actual compiler inputs, extra WAT and options used for the software pixel assertions. owned-new-files.patch preserves this new note; Git remains the source archive.

Retained actual module SHA-256
focused-harness.wasm 656516435a801d7e6a3e90f6538f4828ba3e1918349c60a099c13ca42a418c49
wine-assembly.wasm e41511fe861363fbe1639d80f20cbcdd7240f46937c93a801eaa256ca094c467
wine-assembly.compat.wasm 04acbb89f66ad9814836b5e2c34de659b518f4e1994f50c4ee8069c991f02f72

The shipped hashes match both the locked build and the modules recorded by the native16 regression. Evidence includes logs, module bytes, source patches, native pixel results and checksummed inventory, without a source/media archive. The regression's existing test/binaries/calc.exe input is referenced by its SHA-256 in native16/result.json; it was not copied. Free disk stayed above 2 GB at recorded checks. Browser/server cleanup completed locally; no box ownership remains. The run's top-level result.json is written last.

Next: no remaining work for this fixture repair. The historical Globe band and aligned-reference investigation described in graphics-dx-samples-20261107.md remain outside this change.