GL-to-D3D transform and lighting descriptor checks — 2026-11-01
Both original fixtures passed with their assertions unchanged on main
48951da661440ca501da1a4c20ea414928d257a0. Evidence is retained in
scratch/runs/20261101-gl-descriptor-transform-lighting/ (ignored).
| Original fixture | Result | Wall time | Limit |
|---|---|---|---|
test/test-gl-dfx1-transform.js |
PASS, exit 0 | 3.246 s | 120 s |
test/test-gl-dlt1-lighting.js |
PASS, exit 0 | 3.356 s | 120 s |
These checks execute the real WAT descriptor producers through the existing render harness. They establish the fixture contracts for descriptor production; they do not validate rendered pixels, gameplay, a GPU path, or the full 27-app goal. Neither run failed, trapped, or timed out. No broader crash-free claim is made.
The transform fixture preserves its bitwise matrix transpose checks using non-self-transposing modelview and projection matrices, places modelview in world and identity in view, and builds while projection remains selected. It checks viewport/depth fields, ABI/reserved fields, guard bytes, attribute-stack overflow latching untrusted state, refusal without changing the descriptor, clearing the latch, and null-destination refusal.
The lighting fixture preserves direction negation and rotated eye-space position checks, default material/header values, packed light order, empty-mask success, and material/light specular predicate precision including masked-out lights. Positional lights, excess mask bits, and untrusted state refuse without changing any byte of the caller's buffer. Null destination, latch recovery, reserved words, and both buffer guards remain asserted.
The dispatch baseline was ebed72ce2055025648297c367d116a9aba8256f9.
Work stayed scratch-only until Rayman2 worker
01a1965b-2e10-73b1-92da-1f4523d36d43 supplied fresh final.txt and
result.json in its assigned output directory and explicitly released
runtime/source/build on the board at 09:32:26 UTC. The gate was recorded at
09:33:03 UTC; only then was actual main pinned. Runtime deadline was 09:53:03
UTC and overall deadline 10:55 UTC. The coordinator fallback was 09:39 UTC,
then at least ten-minute intervals; completion preceded that fallback. No
worker was spawned and the existing cap of two was respected.
No retained accepted result for either requested fixture was found in the
scratch result/verification/final evidence search. The accepted canonical build
from d86935e8274cd5d70a53bcc50c5d8bf5632d95c3 was reused: all 4,392
accepted input hashes and 141 accepted actual-read hashes matched, together
with the module and generated region mirror. accepted-build.json and the
retained accepted build log record this; no new build was necessary.
Canonical module SHA-256:
dd43e4a6624d2449d7133b731d687c32530f7fcf0041a69ba71a09f78c09aaf4.
Both fixtures actually instantiated SHA-256
d73eec82b958965566a41de7e4acd4d8990922cf2f727f19f8b41de4924a592a.
The latter is 2,026,958 bytes. All 12,982 production function bodies, every
noncustom section, and all exports match exactly; no extra functions or exports
were added. The canonical module's 37-byte custom section is absent from the
test module, so the modules are not byte-identical.
The reused read-only observer passed actual synchronous and asynchronous Wasm execution against whole 512 MB shared memory hashes, unchanged host/native clocks, and the emitted late-10,001 retention check. It captured each fixture's 228 unique actual file/hash reads, including all 130 source files and bounded descriptor reads from the existing render-artifact snapshot code. Descriptor read coverage includes the manifest, GL matrix producer, D3D fixed consumer, and artifact helper. Each actual read still matched disk after execution. Before/after inventories of 4,444 files were unchanged, as were both original fixture byte copies. Raw compressed stdout/stderr, timeout and Node exit records, actual modules, read hashes, and checks remain in the run directory.
Startup free disk was 2,230,800,384 bytes, satisfying the 2.2 GB requirement. During the scratch-only predecessor wait, externally consumed space fell to 1,361,637,376 bytes; no fixture was launched and no foreign files were cleaned. Space recovered above 2 GB before the gate; fixture checks ended with over 2.25 GB free. No source, test, compiler, helper, configuration, backend, or registration files were edited. No browser, GPU, boat, download, deployment, new EULA, access retry, held-algorithm work, successor, or foreign WIP action was performed. Native local box ID/state/expiry are N/A; local cleanup was delegated to this worker. Cleanup and publication records accompany the run.