Device7 supplied FLOAT3/4 texture transforms

Implemented on 2026-11-04. Evidence is retained in scratch/runs/20261104-legacy-transformed-float34/; the test is test/test-legacy-transformed-float34.js. This closes the supplied FLOAT3/4 immediate/indexed transform gap identified in legacy-coordinate-dimensions-20261104.md, including queued draws. It does not advertise PROJECTED in capabilities or qualify the remaining 27-app scope.

Demonstrated defect and implementation

The predecessor actual WASM, SHA-256 7d096ed4ae107f8a062fb1ffd725b80ef6a0aab9f29f6d8e5764fe9fe45438e9, returned 0x80070057 for the first public FLOAT3 ordinary COUNT2 draw. predecessor-exec.json and its retained log record that public COM execution; the baseline is not inferred from source or a mock.

Device7 now extends its canonical TL input to 40 bytes when an untransformed, supplied-coordinate draw enables a texture transform. The existing 24-byte position/color header is followed by four raw texture floats. FLOAT1/2 retain their existing missing-component defaults, FLOAT3 supplies z and defaults w to one, and FLOAT4 supplies its actual w. The FVF decoder still validates mixed strides and selectors, and the extension reads z/w from the selected original set. Existing public legacy methods and their 32-byte layouts remain in use.

Internal type 5 carries projected inputs; type 6 carries ordinary transformed inputs. Software, the command stream and GPU adapter understand their 40-byte stride. Indexed draws expand the selected full records before using the shared draw path. Clipping copies/interpolates all four raw components, using bounded 400-byte extended polygon scratch rather than overlapping the old state scratch.

Transforms consume the entire supplied vector. Ordinary COUNT2/3/4 retain ordinary perspective interpolation with texture q=1. PROJECTED COUNT2/3/4 select the final requested output as divisor: y, z or w respectively. Both numerators and that divisor reach interpolation; there is no per-vertex divide. Geometric reciprocal w remains separate. A queued record owns its vertex bytes, state and texture matrix. Projected matrix capture normalizes only the selected divisor column; ordinary capture preserves the original matrix. Later matrix, selector or flags mutations cannot change the pending draw.

Execution and independent checks

The authored oracle evaluates full four-component matrix multiplication, geometric w, rational interpolation and an analytic mip/color equation. The separate Python Fraction audit checks recorded pixels against exact rational inputs, rather than calling production texture-transform code. Cases include nonzero z/w matrix rows, distinct divisors, two mixed layouts, selectors 0/1/2/3/7, ordinary and projected COUNT2/3/4, and mip levels 0/2/3.

Execution Result
Direct WebGL1/2, immediate/indexed (four cells) 576 draws, 1,728 pixels
Owning Worker WebGL1/2, immediate/indexed (four cells) 1,152 draws, 3,236 pixels; 288 held pairs
Actual WASM software, immediate/indexed, normal/physical native queue 576 draws, 1,728 pixels
FLOAT1/2 disabled/projected regression 504 pixel checks
Successful-draw semantic negative browser controls 44 controls, 48 draws, 64 pixel observations

All passed. Browser positives also reject 16 invalid selector/count calls. Worker positives require the requested WebGL version, zero fallback/software completion and zero backend errors. Direct cells check public Lock readback and direct GL pixels with GL error zero. Physical queue inspection verifies owned 40-byte z/w records and captured matrices after live-state mutation. Five software negative controls passed as well.

Negative controls submit successful public draws but deliberately drop z, drop w, use the wrong FLOAT3 default, choose the wrong set or change the divisor. Their independently expected pixels disagree as intended. Four additional captured-matrix controls prove that replacing the held matrix changes pixels. The final audit is audit.json: 1,728 positive browser draws and 4,964 pixel observations. Raw browser JSON SHA-256 is 6a10966388a438c6a8fb42705ccccc72832960c932313ff00b6f77b07c5aef83; result.json.gz is a lossless retained copy. Two recovered captures were viewed directly: capture-webgl-1-false.png and capture-worker-2-true.png.

One Chrome/151.0.7922.108 session ran on leased bx_6k76zbug, opened 09:53:59.688Z and closed 10:04:36.173Z; its detached driver exited normally at 10:04:42.683Z. There was no browser failure, reload or second browser. At 10:02:20Z the owned Node driver inspector cached an expensive pure controls() lookup, before any browser failure. The exact expression and before/after control-plan SHA are retained in running-controls-cache.json and cache-running-controls.cjs. The control plan remained c231dabd944ec243aa85b9f19635f4d23d306d3d5f153569b3e2b82eee62dd0f. After closure, the committed helper received equivalent memoization; cache-final-proof.json verifies identical fixture, oracle, config and control outputs. No production bytes or browser fixture changed during execution.

Preflight verified the CJS/ESM Puppeteer closure, module/import/cache graph, HTTP bytes/isolation and actual COM bootstrap for both modules before Chrome. The final locked build reproduces the tested modules:

New pixels were executed with the normal module; compatibility bootstrap is not a claim of compatibility-module pixel coverage. Focused points, batching, fallback counters, depthless behavior, command capture and fixed shader checks passed. An initial misplaced interpolation edit failed compilation and was corrected before runtime; subsequent full builds passed. An initial local control-helper invocation used a non-exported function; the corrected isolated helper check passed. These failures are retained in the logs.

Reuse, custody and limits

reused-evidence.json verifies the exact retained predecessor closure (75 artifacts, eight positive cells, 1,024 draws, 2,960 pixels) and captured-matrix closure (44 artifacts, linking the accepted sixteen-matrix/36-fault evidence). These are historical results by exact SHA, not reruns or new-module results. The predecessor raw result SHA is 543bad998c887631aee31557baf3973715b698c7449d7859136c43efedcd4fc8; the captured remaining-fault result SHA is d7c16d717a7d13aaa7671c706b09be8e02227244280394d1462ff53d837d7213.

All 25 remote artifacts were recovered and hash verified before the recovery deadline. The driver and Chrome are absent, HTTP listener closed and all 46 runtime pins unchanged. The leased box is left RUNNING (last API state idle, healthy/hydrated), expiry 2026-11-04T10:20:16.584Z; shutdown ownership returns to the coordinator. No lifecycle action was taken.

Git SHA plus relevant binary diffs and patches for owned new files preserve source provenance. Modules are referenced by retained path and SHA; no source tree, scratch/docs/test-binary archive or repeated game media was staged. PROJECTED remains unadvertised pending the full relevant contract. Generated coordinates, dimensional vertex buffers/strided/ProcessVertices paths, COUNT1, singular divisors, physical hardware and the broader application matrix remain unqualified by this assignment.