Supplied projected texture: queued and compatibility WebGL completion

The corrected fixture from ee3d94a31c34914d5903c1d966b7dfc2a5e75ed2 now passes its complete rendered matrix. This closes the outstanding browser checks in the prior pixel investigation. No additional production defect was demonstrated; the fixture, production code, independent rational q/w and quotient-derivative mip oracle, tolerance 2, samples and fault controls were unchanged in this continuation. PROJECTED remains clear in the actual public capability descriptors.

Evidence: scratch/runs/20261104-legacy-queued-projected-completion/. The exact tested source commit is 65438de076e5a16458c9bc3b88db7512314b57dd, with relevant uncommitted hunks recorded in relevant-before.diff (empty). source-seal.json records all 42 source, fixture, module and media SHA-256 pins. Git is the source archive.

Rendered results

One Chrome 151.0.7922.108 process ran on assigned bx_9n42uwvf, from 07:36:18.017Z to 07:37:40.613Z, and closed normally. The explicit LEGACY_PROJECTED_HARD_CLOSE was 2026-11-04T07:52:00Z; the 14-minute runtime ceiling was also retained. No repair wait or retry was needed.

Module Route Cells Draws per cell Observations per cell
canonical direct WebGL 4 20 66
canonical owning Worker WebGL 4 40 121
compat direct WebGL 4 20 66
compat owning Worker WebGL 4 40 121

Each group contains direct/indexed public Draw × WebGL1/WebGL2. All 16 positive cells, 480 Draws and 1,496 observations pass, including 136 clipped-away Clear samples. Every observation uses synchronous public Surface7 Lock; the direct cells additionally compare actual GL readPixels. Worker observations require the real legacy WebGL endpoint, zero errors/fallbacks and zero native/fallback software Draw completions.

Each Worker cell includes ten held production batches: capture a projected Draw, mutate live TEXTURE0 and disable projection through public COM, then deliver the captured batch and check all old-state samples before rendering and checking the new state. Across eight Worker cells that is 80 held pairs, including independent geometric w, nonconstant texture q, near-clipped Clear pixels and mip scales selecting levels 1–4. The stable interior post-mutation samples and analytic 1/16-pixel displacement checks from the predecessor are now rendered successfully. Every original projected and clipped sample stays.

The five semantic controls each reject with a rendered PIXEL ORACLE mismatch on canonical WebGL1 and WebGL2: ten rendered rejections total. They preserve successful public Draw and GPU staging. The live-matrix control uses the owning Worker; the other four use direct WebGL.

Fault Actual RGB at rejection Expected RGB
ignore projection 60,70,30 20,70,65
divide before interpolation 20,70,65 60,70,30
suppress mip selection 20,25,30 59,101,92
omit near clipping 60,70,30 11,22,33
consume live matrix after capture 20,25,30 60,70,30

Both versions produce the listed mismatches. counts.json records each cell and rejection. The closed result.json retains public call results, synchronous RGBA arrays, endpoint statistics and injection hashes.

Preflight, recovery and limits

Before Chrome, actual Puppeteer 24.31.0 CJS/ESM imports, retained dependency lock, all staged file hashes, served source/module hashes and HTTP isolation passed. Both actual modules passed public COM bootstrap and public Draw → production adapter/compiler recording transport preflight. Existing module and calc media paths matched their pins; no media or module was copied. Only one small changed dependency, lib/deusex-draw-observer.js, was refreshed to keep the retained Worker dependency graph pinned to the tested commit.

After normal closure all 42 pins were unchanged, Chrome was absent and the fixture HTTP port was closed. closed-process-pins.json records this check. remote-artifacts.json and recover.log verify sizes and full SHA-256 for the unique closed artifacts. Evidence contains exact Git SHA, relevant binary diff and the new note patch, with no source tree or media archive. The four recovered closed artifacts total 126,930,956 bytes; result.json alone is 126,917,519 bytes. A local audit checks every positive RGB against the unchanged tolerance, its position in the retained 32×32 Lock array, direct GL readback where available, Worker counters and all ten negative mismatches. Three selected queued Lock arrays are encoded as queued-readback-0.png through queued-readback-2.png and were personally inspected. These are exact readback images, not canvas-export screenshots. The global evidence checker reports unrelated existing containment errors; none names this run.

Focused depthless-Z, GPU batching, fallback counters, immutable command capture, fixed TL plan cache, fixed-function shaders and retained fill-displacement diagnostic pass. --plan passes its projection compiler and analytic margin checks. flock /tmp/wine-build.lock bash tools/build.sh passes. No local browser, second browser, game, deployment, lifecycle operation or subworker was used.

CDP identifies ANGLE/Vulkan SwiftShader (SwANGLE 5.0.0). Context unmasked strings report Intel UHD 620/Mesa, as in the predecessor; the conflicting identities are retained and do not support a physical GPU claim. Negative controls cover canonical direct draws and the canonical queued live-matrix path; compat and indexed fault variants were not run. Mixed programmed-pixel GPU linkage, COUNT4, generated coordinates, other selectors/dimensions, singular/nonfinite q, wrap/fog combinations, physical hardware and games remain unqualified.

The assigned box is left RUNNING / API idle, healthy and hydrated, expiry 2026-11-04T08:04:15.847Z. Browser, driver and fixture server are closed. Shutdown custody returns to the coordinator. No further runtime is needed for this requested matrix; broader capability qualification is outside this task.