Legacy advertised texture operations through browser WebGL

Passed. Run: scratch/runs/20261106-legacy-operation-webgl/. Tested main: 22d1526cf7b249d40151ed3419484f5380ac4b22, plus the recorded new-file patch for this regression. No production defect was demonstrated; no emulator implementation or advertised capability was changed.

The new focused test/test-legacy-texture-operation-web.js reuses the public x86 COM fixture in test-legacy-texture-rendered-contract.js. It targets both shipped modules, direct WebGL and the owning CPU/render Worker route, and explicit WebGL 1 and 2 contexts. Native backend topology is not qualified.

The argument matrix covers all 144 operation/argument combinations twice: RGB and an independent alpha witness. Binary channel vectors keep all same-operand products exact; eighteen richer independent channel vectors exercise integral nontrivial products. White RGB with SRCALPHA/ZERO blending exposes stage alpha as RGB on the actual X8 render target. It does not infer alpha from the unused X byte. ARGB source texture formats, setter round trips, stdcall stack cleanup and output guards use public COM calls.

CURRENT, DISABLE, all nine unbound operation pairs, peer device isolation, and a held owning Worker batch with changed live setters are included. Public GetTexture verifies binding retention through DISABLE and null while unbound; re-enabling without rebinding restores the texture. DrawPrimitive7 and DrawIndexedPrimitive7 alternate. Operations 5–10 remain explicit UNADVERTISED diagnostics. TextureOpCaps must remain exactly 0xf. Semantic controls change operation/argument lowering or captured Worker state; they must fail real pixel arithmetic after successful draw calls.

Aranna0205, Conan0240, then KKnD0245 final releases and independent process, port, profile, lock and disk checks preceded the immutable execution window: 03:56:06–04:41:06 UTC. All tests finished by 03:58:43. The first cheap dependency check caught an assumed Puppeteer directory layout before Chrome; the runner now imports the installed package's resolved entry. That failed attempt is preserved separately. Public COM preflight then passed 5,557 calls per exact module with no draws, covering resource formats and every case's setters.

Exactly one Chrome 151.0.7922.108 session ran from 03:57:44 to 03:58:42, exiting normally with code 0. Actual context VERSION strings proved WebGL 1 and 2; the reported renderer was ANGLE / Intel UHD Graphics 620 / Mesa 23.2.1. The result does not infer the renderer from requested launch flags. The supervisor also observed a transient timezone-probe profile; both that profile and the test profile were absent after the session closed.

Module Bytes SHA-256
wine-assembly.wasm 2176554 3bc4e29b86d3a085bf14de1a242f87f22321d513f26cb2259f22804d0c32dc7e
wine-assembly.compat.wasm 2179200 473ec05b5b10c5d6fa653d042b0aec4821804d6e6c5adf1e3ec8c0d744d4f2bc

All eight module × direct/owning-Worker × WebGL-version cells passed. Each cell executes 340 operation/argument/edge/diagnostic cases, plus peer and DISABLE/re-enable checks. Direct cells passed 343 draws and 344 public COM readbacks each; Worker cells passed 345 draws and 347 readbacks each. Total: 2,752 positive GPU draws, 2,764 completed 8×8 readbacks, and 530,688 independent RGB channel comparisons across their full targets. Direct cells also check actual gl.readPixels results. All routes report zero GPU errors/fallbacks; owning Worker statistics additionally prove zero native/fallback rasterization and actual legacy WebGL endpoint identity. Each draw has a real fence/readback.

Ten controls (five per WebGL version) reject at their designated pixel observation after successful public draws: swapped SELECTARG lowering, lost TEXTURE arguments, ignored alpha operation, wrong CURRENT mapping, and changed captured Worker operands. The final control edits only the held command's actual 4KB state. Served mutation hashes and captured before/after operands are retained. Expected arithmetic is unchanged by every control.

The restored operation-caps suite passed its 15 rejecting controls, supported pairs, 144 argument cases and six unadvertised diagnostics. DISABLE (three rejecting reversions), DISABLE capability (seven controls), GPU batching and GPU points checks passed. The required locked build passed, including after the dependency-loader repair. Final WAT closure, production JS and module hashes matched the tested inputs. All eight generated positive captures were personally inspected; they show generated regression tiles, not gameplay.

Run after building with an explicit ignored evidence directory and fixed deadline, for example:

LEGACY_OPERATION_EVIDENCE=scratch/runs/EXAMPLE/browser \
LEGACY_OPERATION_HARD_CLOSE=YYYY-MM-DDTHH:MM:SSZ \
node test/test-legacy-texture-operation-web.js

--plan is static only; --preflight exits before Chrome after public COM checks. The runner uses both existing shipped modules, pins their served buffers, checks disk hashes before each route, and rechecks source/module identity afterward. Unavailable runtime dependencies or contexts fail.

Limits: Device7 TL triangle lists only, including indexed public submission. The native render backend's unsupported topology is not presented as browser coverage. Other device generations, selector modifiers, unadvertised texture operations, gameplay and the full27 suite are not qualified. Operations 5–10 are observed diagnostics, including ADD, with TextureOpCaps still exactly 0xf. No boxes, game media copies, source archives, deployment or subworkers were used. Git SHA, relevant binary diff and new-file patches preserve source; the evidence contains logs, JSON and generated captures.