Repaired quotient-gradient projected WebGL matrix

All sixteen positive WebGL cells pass with the repaired analytic projected mip atlas. The fault matrix stopped at a demonstrated captured-matrix runner scope bug, now repaired and tested offline. The complete fault matrix remains unqualified. PROJECTED remains unadvertised. Generated coordinates, other input dimensions, full27 and physical hardware remain unqualified.

Read with the first actual mip failure and offline repair, the expanded supplied-coordinate contract and the historical COUNT3 matrix. The historical fixture is not evidence for the expanded matrix.

Evidence is retained in scratch/runs/20261104-legacy-repaired-quotient-webgl/. The exact source revision, relevant binary diff and path/hash pins are in git-sha.txt, relevant-before.diff and source-seal.json. Git is the source archive. Only changed small backend/shader/runner dependencies were staged; retained media and module paths are referenced by full SHA-256. Both runtime modules already matched the authorized hashes under the build lock, so no module/media copy was necessary. The locked full build reproduced both hashes.

The runner reuses the exported fixtureSource, plan and independent rational oracle from test/test-legacy-projected-coordinate-contract.js, through the existing real browser harness. Tolerance remains two bytes. All three counts, selectors 0/1/2/7, independent geometric W and texture q, analytic mip levels 1–4, near clipping, held selector/matrix/flags mutation and six semantic fault controls remain. Recording/compiler submissions are separate from GL pixels. The runner now identifies the published repair commit and requires an explicit hard-close environment value instead of a stale historical default.

Actual rendered results

One Chrome 151.0.7922.108 process ran on assigned bx_a9w3p8hk from 08:42:16.986Z to 08:49:23.629Z, then closed normally at the first unexpected failure. The explicit hard close was 09:10Z. No second browser or retry ran. Tested source is cfad0fbee91fb0c921d0f72ca0a7832a5890c052, including the runner/other relevant working hunks in relevant-before.diff. It includes the published production quotient repair from a62801c2f653b738a9ba6fa052c91b9d7c673d52.

Module Route Cells Public draws per cell Pixel observations per cell
Normal Direct GL 4 155 465
Normal Owning Worker GL 4 309 866
Compat Direct GL 4 155 465
Compat Owning Worker GL 4 309 866

Each group covers WebGL1/WebGL2 × public direct/indexed Draw. All 3,712 public Draws and 10,648 positive observations pass, including 72 clipped-away Clear samples and 616 held pairs. Each held pair captures the old draw, mutates live selector/matrix/flags through COM before delivery, then checks the old pixels and subsequently rendered new state. Public Surface7 Lock supplies asserted RGB; direct routes also agree exactly with independent GL readPixels. Worker observations require the actual legacy GL endpoint and zero errors, fallback, native-software or fallback-software draw completions. PROJECTED stays clear in public descriptors throughout.

The previously failing COUNT3/TEX7/scale16/wfalse sample (9,11) now reads 138,87,119 in all sixteen cells, exactly matching the independent oracle. The former result was 137,163,146. No further production rendering defect was demonstrated in this run.

First failure and repair

The canonical nonindexed WebGL1 COUNT2 faults for ignored projection, wrong selector, wrong divisor, early division and suppressed LOD each reach a successful public Draw and a rendered PIXEL ORACLE rejection. The first four read 60,25,65 versus expected 20,25,30; suppressed LOD reads 60,25,65 versus 59,56,57. These are deliberate semantic injections, not production defects.

The sixth fault, COUNT2/WebGL1 owning-Worker captured-matrix replacement, failed with ReferenceError: identity is not defined. The delayed delivery callback was outside the scope containing the public setup's identity constant. It therefore failed before the intended corrupted-record pixel rejection. Its earlier uncorrupted draws are not counted as successful fault coverage. All thirty later fault variants were unexecuted; this repaired variant plus those thirty still need actual rendered qualification. The positive matrix remains valid independently of this test-only fault failure.

test/test-legacy-count-selector-webgl.js now injects a self-contained corruptCapturedMatrices helper with a local identity literal. Its regression runs the exact helper in an isolated realm with no identity binding, on framed type-3 and type-4 records, and executes the exact emitted postMessage wrapper through held then drained delivery. Reintroducing the old scope reference in memory makes that regression fail with the original ReferenceError. Only the type-4 owned matrix changes; adjacent record bytes and guards remain exact, and invalid framing rejects. The emitted fixture parses, analytic plan/repair controls pass, and both predecessor production files still fail the quotient regressions. This is an offline runner repair, not a successful GL captured-matrix fault. No production file was changed in this task. relevant-final.diff separates the post-closure repair from the exact browser-tested source.

Validation, recovery and custody

Separate --preflight passed before Chrome: actual Puppeteer 24.31.0 CJS/ESM imports and retained dependency lock, all 44 source/module/media pins, served HTTP hashes/isolation, both exact modules' public COM bootstrap and production recording transport. Browser cache graph checks pass: 98 page / 13 Worker scripts. The retained calc path is test/binaries/calc.exe, SHA-256 cb71fcc9726f756421519a215c39cf89efd930c9ae0fa5bbb59b64050faef52b.

Retained runtime module SHA-256
build/wine-assembly.wasm 129a71a01a5fbc8b07d6fa242a5a3ebc6a29f5aed9fb4b6a47d9318a764c1f52
build/wine-assembly.compat.wasm 3fde0fc42ecd0d9d5096d0b9eb83cd3086cb74aab8ac3f7b9f3fd15f5783fe6a

Both full expanded native software suites pass: four routes with 155 draws / 465 observations per route, actual held native Worker replay, all eighteen rendered fault rejections per module, and separate unqualified generated characterizations. Both public caps suites pass 47 calls / 55 descriptors. Shader, fixed-function shaders, fixed plan cache, depthless-Z, batching, fallback counters and solid transport plan checks pass. The locked full build passes both before Chrome and after the runner repair. No recording submission or software result is counted as GL pixels.

remote-artifacts.json / recover.log verify all five closed remote artifacts by full SHA-256 and size. result.json.gz losslessly recovers 192,034,623 bytes, original SHA-256 da6582bac3cddcde051165dbc46c28f2f3c9a6ce825d39e7509f403cec0f890f. The local audit checks every positive RGB against unchanged tolerance and its 32×32 Lock array, direct GL readback, Worker exclusion counters, all five completed rendered negative controls and the exact first unexpected failure. Three queued COUNT2/3/4 readback PNGs were encoded from retained Lock arrays and personally inspected. No source tree archive, repeated media or duplicate module recovery was made. The global evidence checker reports existing errors; none names this run.

After closure, all 44 remote pins match, Chrome is absent and the fixture server is closed. The runtime retains the exact browser-tested source; the post-closure runner repair is local and published with its diff. Box inspection confirms healthy/hydrated RUNNING / API idle, expiry 2026-11-04T09:20:14.588Z. Shutdown ownership returns to the coordinator. No allocation, resume, stop, source-box access, local browser, subworker, successor, game, Heroes II or deployment was used.

CDP reports ANGLE/Vulkan SwiftShader; context unmasked strings report Intel/Mesa. Their conflicting identity does not qualify physical hardware. Generated coordinates, other input dimensions, full27 and hardware remain unqualified. Next is a separately authorized run of the repaired captured-matrix control and remaining semantic variants, with refreshed source pins/preflight.