Supplied FLOAT3/4 through public legacy vertex buffers

Implemented 2026-11-04. Evidence: scratch/runs/20261104-legacy-dimensional-vertex-buffer/. Regression: test/test-legacy-dimensional-vertex-buffer.js (explicit e2e tier).

First demonstrated defect

Main already contained 4ef1febf63a37350fe3889da5b406084b224672c and its immediate FVF dimensional repair. The actual predecessor module at build/wine-assembly.wasm, SHA-256 c2106403fdac5e30cf52577089fb83c95714623caf8b97a88176a12e97786517, still lost supplied z/w through the public VB route.

baseline-semantic/first-failure.json records actual x86 COM execution: DirectDrawCreateEx, QI IDirect3D7, CreateVertexBuffer7, Lock/Unlock and DrawPrimitiveVB7, dereferencing the published vtables. The first mixed-layout FLOAT3 set 2, ordinary COUNT2 draw returned S_OK with correct stdcall cleanup. Pixel (5,5) was RGB [60,70,30]; the independently authored matrix, interpolation and texture equation expected [60,25,65]. This is the exact accepted direct-path module, not a reconstructed predecessor.

Both VB draw functions packed the selected coordinates into a 32-byte record, discarding z/w, and bypassed the immediate path's 40-byte supplied-vector conversion. ProcessVertices was not needed to expose this first defect.

Repair and regression

The two shared VB draw functions now validate the supplied selector/FVF and reuse d3dim_d7_supplied_type / d3dim_d7_supplied_vertices for untransformed draws with enabled texture transforms. The original source pointer includes the start-vertex offset. Existing clamping, indexed scan, lighting, expansion, queue ownership, matrix capture and rasterization remain the shared machinery. There is no new rendering representation or capability advertisement.

The regression reuses the exact accepted fixture setup and analytic oracle from test-legacy-transformed-float34.js, replacing submissions with public VB calls. All buffers are created, locked and unlocked through COM. Two poisoned records precede the start offset of two. Indexed cases store v1,poison,v2,v0 and fetch indices 3,0,2. Queued indexed cases pass 0xffffffff as the vertex count to exercise capacity clamping and the highest-index scan.

Final execution Draws Pixel assertions
FLOAT3/4 software, ordinary/indexed 288 864
FLOAT3/4 physical native Worker queue, ordinary/indexed 288 864
FLOAT1/2 defaults and disabled FLOAT1–4 transforms, ordinary/indexed 184 552
Total 760 2,280

Coverage includes two mixed dimensional layouts, selected sets 0/1/2/3/7, ordinary and projected COUNT2/3/4, varying geometric w, and mip levels 0/2/3. The physical queue is held before delivery. Public VB Lock/Unlock overwrites its live bytes, index bytes are overwritten, and selector/flags/matrix change before replay. All 288 captured records retain all four selected components, the FLOAT3 w=1 default, the captured matrix and the expected pixels.

Each of six positive cells also checks eight no-draw guard calls (past-end, zero count, null VB, null indices and out-of-range index), all 1,024 unchanged clear pixels, the unchanged full VB and two E_INVALIDARG selector/count cases. Six successful public VB draws with deliberately wrong selector, z, w, missing-w default, matrix component or indices fail the unchanged pixel oracle.

audit.py independently applies the accepted Python Fraction equation to all 2,280 final pixels. It verifies exact SHA/size of the accepted direct evidence and fixture rather than rerunning or claiming new direct/GPU coverage. The accepted compressed result SHA is 4627d5de45b758d486f4b1ca7a358844b1f2590a44279c8e875e6dacc812ac9c; its uncompressed SHA is 6a10966388a438c6a8fb42705ccccc72832960c932313ff00b6f77b07c5aef83. The original direct evidence remains at scratch/runs/20261104-legacy-transformed-float34/.

Checks and limits

Two serialized flock /tmp/wine-build.lock bash tools/build.sh runs passed. The final normal module SHA is 0b99df572d10ad7d2ca4c3d88dbe0e3b03cc24b607a9481654e08570d51e6b20. test-d3dim-v3-vertex-buffer-draw.js passes (900 indexed pixels), as does test-d3d-legacy-vertex-matrix.js. Tier membership and owned diff checks pass.

test-legacy-coordinate-dimensions.js --offline reaches its historical assertion that transformed FLOAT3/4 must return E_INVALIDARG and fails because the accepted immediate repair correctly returns S_OK. Its positive pixels precede that assertion; this is not claimed as a passing suite. The owned new test uses the accepted successor oracle and tests FLOAT1/2 plus disabled FLOAT3/4 through VB. The historical file was not edited. Logs also preserve an initial fixture slot-number error, a missing output-directory error, and an audit filename mismatch, all corrected before final validation.

No Chrome, remote box, lifecycle action, deployment or game media was used. No new compatibility-module pixels or WebGL pixels are claimed. The shared Device3 VB entrypoints retain their existing focused regression, but the new dimensional COM cells use Device7. ProcessVertices, strided/generated coordinates, COUNT1, singular divisors, physical hardware and application coverage remain unqualified. In particular the ProcessVertices producer still needs its own dimensional/multiple-set investigation; this repair does not claim to preserve those producer outputs. PROJECTED, cube and BORDER remain unadvertised.

Git SHA, owned binary diffs/new-file patches, dependency/module path hashes, losslessly compressed closed logs and an artifact seal are retained in the run. No source tree or repeated module/media snapshot is included.