Public ProcessVertices dimensional texture-set preservation

Implemented 2026-11-04, starting at main 2e958d0dc7cc5e76df0364532979f19a1da20604. Evidence: scratch/runs/20261104-legacy-processvertices-dimensional/. Public regression: test/test-legacy-processvertices-dimensional.js (e2e tier).

Demonstrated predecessor failure

The actual accepted normal module, SHA-256 0b99df572d10ad7d2ca4c3d88dbe0e3b03cc24b607a9481654e08570d51e6b20, was executed before compiling the repair. Public DirectDrawCreateEx, QI, CreateVertexBuffer7, Lock/Unlock, ProcessVertices7 and VB draw calls dereference the published COM vtables and execute their x86 thunks. Each call checks stack cleanup. Source FVF includes XYZ/point-size/diffuse/specular; destination FVF includes XYZRHW/diffuse, with independently laid-out mixed texture dimensions.

baseline/first-failure.json records the first byte mismatch: destination byte 324 was 0 instead of 49. baseline-pixel.json records a separate draw with the byte assertion bypassed to observe its consequence. The successful public FLOAT3 set-2 draw at (5,6) produced RGB [20,70,65], versus independently expected [60,25,65]. Both failures name the accepted module SHA. Expected bytes and the analytic pixel equation were authored before the producer edit.

ProcessVertices packed one selected UV pair into a canonical 32-byte vertex, then replicated that pair into every output set at fixed eight-byte intervals. This lost unrelated sets and FLOAT3/4 components; FLOAT1 also made those writes land at incorrect offsets. The accepted VB draw fix alone could not recover the destroyed producer output.

Bounded producer repair

Position, normals, lighting, diffuse and specular still use the existing canonical pack/prepare/unpack path. Texture output now copies each original set separately, using source and destination FVF component counts and offsets. Destination components absent from the source use the established supplied vector defaults, zero for x/y/z and one for w. Truncation does not read beyond the source set. Invalid texture-set counts above eight return E_INVALIDARG before conversion. Generated-coordinate selectors retain the existing canonical generated UV route; there is no new generation capability.

The existing source/destination capacity clamps precede all reads and writes. When both objects use the same allocation, a temporary copy preserves the original source records across overlapping output ranges. Allocation failures release temporary storage. The shared producer serves both legacy and v7 entrypoints. No draw representation, capability advertisement, PROJECTED, cube-map or BORDER support was added.

Verification

final-complete/ contains four actual normal-module software cells: ordinary and indexed VB draws, each immediate and through the physical native Worker queue. Each cell draws 23 states: two mixed layouts, selected sets 0/1/2/3/7, FLOAT1–4, constant/varying geometric w, differing destination dimensions, and missing source sets. Source index two and destination index three are nonzero. Both source-limited and destination-limited calls pass count 0xffffffff. Indexed draws fetch 3/0/2 from a pool with a deliberately distinct extra record.

audit.py independently lays out all destination bytes with Python struct packing and checks all pixels with the accepted exact Fraction equation. It also verifies public call arguments/results/stack cleanup and exact SHA/size of seven accepted direct/VB artifacts, including the accepted direct raw result SHA 6a10966388a438c6a8fb42705ccccc72832960c932313ff00b6f77b07c5aef83. Those historical direct/browser/GPU results are reused by hash, not rerun or claimed for this module. An initial audit included a bootstrap record without stack fields; restricting that assertion to recorded stack-bearing calls corrected the audit. Its failed log is retained.

Two full serialized flock /tmp/wine-build.lock bash tools/build.sh runs pass. The final normal module SHA is b748d9b5a16631762be84a6e1017ca095c192623bc288657488157d32f79c0ca; compatibility build SHA is cafdf1fd380aadb73571c68fbd77612ba91d58178a05a6afc469ae858394c333. Focused existing tests pass: test-d3dim-process-vertices.js (both stdcall aliases, position/color/UV/clamps), test-d3dim-process-vertices-fog.js, test-d3dim-texgen-camera-position.js, and test-d3dim-v3-vertex-buffer-draw.js (900 indexed pixels).

Limits and evidence recovery

These public dimensional cells use supplied coordinates, disabled texture transforms and XYZRHW output. FLOAT3/4 components are checked in public Lock readback; the subsequent TL draw and its immutable queue record consume UV. This does not establish a dimensional texture-transform contract for already transformed ProcessVertices output. Generated-dimensional outputs, strided inputs, physical GPU/application coverage and broader ProcessVertices flags remain unqualified. The next gap is that transformed-output texture-transform contract. Existing generated UV arithmetic has its focused regression; no new generated pixel result is claimed here.

The historical coordinate-dimensions test is imported only for its immutable plan. Its stale FLOAT3/4 rejection postlude was neither needed nor edited. No Chrome, remote allocation/runtime/lifecycle action, box, game media, deployment, worker or successor was used or assigned.

Git commit SHA plus relevant git diff --binary and small new-file patches preserve source provenance. Module/dependency paths and hashes identify inputs; rebuild from the recorded Git revision and patch with tools/build.sh to recover module bytes. Closed large logs/JSON are gzip compressed with verified roundtrip SHA; the compression index and artifact seal preserve exact recovery. No source archive, repeated module or media copy was made.