Software GL raster and multitexture — 2026-11-01

Both original, unchanged fixtures passed against actual instantiated WAT and software DIB pixels. No assertion failed; neither invocation crashed or trapped.

Fixture Result Elapsed
test/test-gl-software-raster.js PASS, exit 0 3.877 s
test/test-gl-software-multitexture.js PASS, exit 0 3.804 s

Raster retained the original asymmetric geometry, Y direction, translated modelview/projection order with PROJECTION selected, both windings with default culling disabled, explicit back-face culling and polygon attribute restoration, and near-plane rejection/clipping assertions. The stored opaque red sample was 0xFFFF0000; the opposite winding painted more than 150 green pixels. The near-straddling floor sample at (VP/2, 3*VP/4) retained RGB 0x00FF00. The fixture also retained its original sampler, texture, depth, alpha, scissor, blend, fog, lighting and far-plane assertions. No assertion was weakened or geometry replaced. The fog assertions inside this unchanged raster fixture ran; the separate lines/fog fixtures were not rerun.

Multitexture retained independent unit-1 coordinates, MODULATE and REPLACE, enable push/pop, alpha testing after unit 1, border coordinates and channel bounds, per-texture border state, invalid-sampler preservation and deletion reset. Exact left/right controls include pink 0xFFFF8080 on both halves with unit 1 disabled, pink/dark blue 0xFF000080 under MODULATE, and white 0xFFFFFFFF/blue 0xFF0000FF under REPLACE. Four queued triangles were replayed through the native worker entry point: the original assertions require captured red-border 0x7FFF7F7F and blue-border 0x7F7F7FFF pixels on both halves despite later setters. These assertions all passed.

Gate, prior results and build identity

Evidence: scratch/runs/20261101-gl-software-raster-multitexture/ (ignored). Preparation before the gate was scratch only, with no tests, builds or note edits. DarkReign2 worker 01a19615-8d76-7902-973e-211aa0130a2b supplied fresh worker final.txt/result.json at 08:07:25 UTC and an explicit 08:07:19 runtime/source/build RELEASE. The lane was acquired at 08:07:51 UTC; baseline current main was d9aca1952251236c2090144ffad7ca63977a8c10.

The lines/fog and texture-sweep notes were read first. Searching existing run-result, verification, final and reverse-engineering notes found an older raster exit-0 report in 20261031T0443Z-land-seven-topics, but no accepted matching-current-input identities for that result, and no accepted exact multitexture result. It was therefore insufficient for reuse. Mipmap, texgen, lines and the separate fog fixture were not rerun.

The accepted locked build from 20261101-dark-reign2-native-mip-implementation was reused: all 4,392 accepted source hashes and 141 accepted actual-read hashes matched, together with the production module and generated mirror. All 4,444 files in this run's broader input snapshot remained unchanged. No build was needed and no source, fixture, compiler or repository helper was edited.

Actual modules and execution limits

The actual raster module is 2,027,021 bytes, SHA-256 9857210e2104c84e713cd28cfd6d9f721f75d8e5efac42cc7657de92729cbf0a. Its only differing non-custom section is exports: the original fixture adds border_sample_test, key_sample_test and decal_combine_test.

The actual multitexture module is 2,027,043 bytes, SHA-256 afd0d7de118bd88f6d234e9eda8d686f987210abbbe30481d2b5244b8b07b107. Its function, export and code sections differ because the original fixture adds border_set_test, border_ptr_test, tex_slot_test and the queue_enable_test function, which sets gl_sw_worker_ok to zero for the original deterministic queue replay. Every production function body remains byte-identical in both modules; all production exports retain their indices. Both actual modules omit the production build's 37 bytes of custom metadata. Neither actual module is byte-equal to production.

The original entrypoints ran sequentially on local Node v24.18.1 under timeout --signal=TERM --kill-after=5s 120s, using the byte-identical accepted read-only forward-only observer from the lines/fog run. It retained actual instantiation bytes and 88 unchanged file reads per fixture. The unchanged render helper compiled the WAT closure with each fixture's original extra WAT, instantiated shared 512 MB memory, and the assertions read software DIB pixels.

Evidence includes full original fixture copies (exact assertions), compressed original stdout/stderr, exit codes and elapsed times, actual modules, loaded reads, before/after manifests, accepted build log, gate, verification, cleanup, publication, final.txt and result.json. Own test PIDs/groups and watcher were checked absent; no temporary file or box was owned (ID/state/expiry N/A). Free disk stayed above 2 GB. This establishes these two bounded fixtures only: no gameplay, GPU, full27 or general crash-free claim follows. Held algorithms were not exercised separately. No further runtime work is required here.