Supplied FVF coordinate dimensions: bounded audit and packing repair
The new audit demonstrates a production FVF layout defect and repairs it. Immediate Device7 draws now decode mixed FLOAT1/2/3/4 texture-set layouts, including selector offsets and the FLOAT1 missing-v default. Eight new normal module WebGL cells pass. The sole browser run was terminated by an incorrectly short launcher timeout after those positives; new WebGL negative controls remain unexecuted. Their software pixel controls pass. No second Chrome, reload or continuation of the browser audit was attempted.
The qualified boundary is deliberately limited:
| Supplied input | Qualified immediate DrawPrimitive7 / DrawIndexedPrimitive7 behavior |
|---|---|
| FLOAT1 | Disabled transform samples (u,0); projected COUNT2/3/4 transforms (u,0,0,1) |
| FLOAT2 | Disabled transform samples (u,v); projected COUNT2/3/4 transforms (u,v,0,1) |
| FLOAT3/4 | Disabled transform samples the first two supplied components, ignoring extra components for this 2D sampler |
| FLOAT3/4 with enabled texture transform | Explicit 0x80070057 rejection; extra input components cannot be preserved by the current canonical record |
| Supplied selector outside a nonempty FVF's declared sets, or more than eight sets | Explicit 0x80070057 rejection |
This is not full transformed FLOAT3/4 support. Ordinary nonprojected COUNT transforms, projected singular divisors, dimensional VB/strided/ProcessVertices rendering, generated coordinates, full27, physical hardware and gameplay remain unqualified. PROJECTED is still unadvertised. No capability was widened.
Existing contract and demonstrated defect
Read with the accepted matrix/fault audit.
Its expanded native fixture used FVF 0x8e2, with eight FLOAT2 sets and no
TEXCOORDSIZE bits. It did not exercise dimensional FVF layouts. The accepted
sixteen-cell browser matrix and 36 semantic controls were not rerun in Chrome.
Production fill_primcaps advertises texture caps 0xF3DF with PROJECTED
0x400 clear. Device7's public dwFVFCaps is eight texture sets; every new
cell's captured GetCaps confirms it. The count selects an output dimension,
while TEXCOORDINDEX selects a supplied set. Neither changes the storage layout
specified by the FVF. The D3D7 TEXCOORDSIZE encoding is FLOAT2/3/4/1 = 0/1/2/3
in successive two-bit fields beginning at bit 16.
The old d3dim_fvf_stride always added eight bytes per set. All three general
packing branches also skipped eight bytes per preceding set and unconditionally
read a second float. On the retained old normal module, the new first FLOAT1
public Draw returned success, but sample (5,5) remained Clear [11,22,33]
instead of [20,25,30]. predecessor-exec.json preserves the actual-WASM
failure before the old module was replaced. This was a real pixel mismatch,
not a parser or setup exception.
The repair sums each encoded dimension for stride and selector offset. FLOAT1 packing leaves the zero-initialized second component intact. The exact FLOAT2-only MW3 fast path remains valid. Device7 immediate handlers reject unsupported transformed FLOAT3/4 and invalid supplied set/count boundaries before packing, with the correct stdcall stack cleanup. Existing other methods are not included in this new validation claim. The shared stride change also affects VB allocation; existing VB capacity, packing/unpacking, alias, lifetime and stack regressions pass, but dimensional VB rendering is not qualified.
New pixels and controls
test/test-legacy-coordinate-dimensions.js authors 92 new states with mixed
layouts 12341342 and 43214231, selectors 0/1/2/3/7, independent geometric w,
and projected COUNT2/3/4 for FLOAT1/2. Both layouts contain twenty coordinate
floats per vertex; their selected-set offsets differ. Nonzero unused matrix
rows and translation distinguish the (…,0,1) input defaults. COUNT2 uses y
as divisor, COUNT3 z, and COUNT4 w; the two-dimensional sampler uses the first
two quotients. The new matrix has distinct divisor columns. Higher input
components are truncated only in the tested disabled-transform case.
Expected pixels come from authored coordinates and a standalone rational
interpolation / analytic mip oracle. No production FVF decoder, adapter,
sampler, shader or captured matrix supplies expected values. Tolerance stays
two bytes. audit.py independently recomputes all 2,960 browser observations
with Python Fractions and the authored color equation. It verifies public Lock
bytes, direct GL readPixels/error, all captured caps, and owning Worker backend,
context version and zero software/fallback/error counters. Observed mip levels
are 0, 2 and 3; this run does not extend the prior full mip-level qualification.
| Route, for each WebGL version 1 and 2 | Cells | Draws per cell | Pixels per cell | Expected rejections per cell |
|---|---|---|---|---|
| Direct, nonindexed/indexed | 2 | 92 | 276 | 14 |
| Owning Worker, nonindexed/indexed | 2 | 164 | 464 | 14 |
Totals: eight cells, 1,024 successful Draws, 2,960 observations, 144 held matrix/selector mutation pairs and 112 explicit HRESULT rejections. The normal module supplies these pixels. Both exact modules pass public COM bootstrap; no new compatibility-module pixel claim is made.
Separate actual-WASM software execution passes both direct/indexed positives (552 observations and 28 rejections) and five semantic controls. Every control reaches a successful public Draw and rejects non-Clear rendered pixels at the pixel oracle: wrong selector, nonzero missing v, missing default w, wrong COUNT divisor, and swapped FVF dimension fields with unchanged total stride. These controls do not earn credit through setup failures. The same five controls are planned for direct and owning Worker WebGL1/2, but none completed before the launcher failure. The prior 36 controls remain their previously accepted scope.
Browser closure and tested launcher repair
The sole Chrome 151.0.7922.71 on bx_gyunxs8u opened at
09:27:04.295Z. All eight positive rows were saved. boat exec --timeout 60
then terminated the process at 09:28:03.202Z, exit 143 / SIGTERM, before any
new negative result was saved. Raw result.json therefore correctly retains
PREPARING, without a fabricated PASS or normal-close timestamp. The first
closure inspection at 09:28:17.551Z finds both Chrome and Node absent and
the HTTP port closed. No rendering semantic failure was observed in completed
positives. No browser was relaunched.
The concrete invocation fix is to use boat exec BOX --detach -- COMMAND, then
poll boat exec BOX --status PID. --detach cannot be combined with --timeout.
An actual detached Node control ran for 65.442 seconds and exited zero, beyond
the failed 60-second boundary; launcher-fix-start.json and
launcher-fix-status.json preserve its complete output and timestamps. This
validates the launcher fix without consuming a second browser. The fixture's
own 14-minute timer and explicit hard close remain the browser limits for any
separately authorized future invocation.
Recovery, validation and custody
Evidence: scratch/runs/20261104-legacy-coordinate-dimensions/.
The actual tested checkout SHA and production hunks are in git-sha.txt,
source-seal.json and relevant-before.diff; the owned new test is preserved
as owned-test-before.patch. The final test changes only a layout comment;
comment-only-proof.json verifies identical serialized fixture, oracle and
configuration. source-provenance.json records source hashes.
Git is the source archive. No source archives, media copies, module evidence
copies, allocation, source-box access, local heavy browser, subworker,
successor, game or deployment was used.
The required locked full build passes and reproduces both transferred modules:
| Module | SHA-256 |
|---|---|
| Normal | 7d096ed4ae107f8a062fb1ffd725b80ef6a0aab9f29f6d8e5764fe9fe45438e9 |
| Compat | 51c0ab6770512fc5b2f1aa4b561edcb6910630e47628303de3edfc60b3c6a101 |
Modules were transferred directly into runtime build paths under the build lock, with full SHA verification. Before Chrome, local cache graph checks (98 page / 13 Worker scripts), generated-region build gates, actual Puppeteer CJS/ESM imports, dependency lock, all 45 runtime pins, HTTP bytes/isolation, public COM bootstrap of both modules, and offline positive/negative tests pass. The existing native expanded fixture's software/physical Worker tests and the FVF lighting / VB regressions also pass. The slim runtime has no full index page or cache checker; its actual imported Worker closure is pinned. Some initial local helper/command mistakes are preserved as command failures, not pixels.
remote-artifacts.json and recover.log verify fourteen recovered artifacts,
including all eight screenshots. Direct WebGL1/nonindexed and owning Worker
WebGL2/indexed captures were personally viewed: textured triangles and clear
outside regions are present. They do not qualify hardware or gameplay. CDP identifies SwiftShader / SwANGLE.
result.json.gz losslessly retains the original 53,140,510-byte partial
result, SHA-256
543bad998c887631aee31557baf3973715b698c7449d7859136c43efedcd4fc8.
closed-process-pins.json confirms all 45 pins unchanged, Chrome absent and
HTTP closed. Large local software/regression JSON is retained losslessly with
original and compressed hashes in lossless-compression.json.
Custody returns to the coordinator: bx_gyunxs8u remains RUNNING / API idle, healthy and hydrated, expiry 2026-11-04T09:56:26.965Z. Shutdown ownership remains with the coordinator. No box lifecycle action was taken. Remaining work is transformed FLOAT3/4 support and the unexecuted new WebGL controls; this worker stops after publishing the bounded repair and exact first failure.