Captured-matrix and remaining rendered semantic faults

All 31 remaining semantic fault variants pass actual rendered rejection, including the repaired captured-matrix callback. Combined with the five previously accepted rejections, all 36 controls are covered. No further production rendering defect was demonstrated. PROJECTED remains unadvertised; generated coordinates, other input dimensions, full27 and physical hardware remain unqualified.

Read with the accepted positive matrix and callback repair. Evidence is in scratch/runs/20261104-legacy-captured-matrix-faults/.

test/test-legacy-count-selector-webgl.js --remaining-faults skips the accepted positive matrix and its first five completed negative controls, beginning at COUNT2/WebGL1 owning-Worker captured-matrix replacement. Its offline controls assert the exact 31-case suffix and the unchanged default 36-case selection. The default invocation still runs all sixteen positive cells and all faults. The independent rational/mip oracle, two-byte tolerance, counts 2/3/4, selector coverage and held selector/matrix/flags mutation are unchanged.

Actual run

One Chrome 151.0.7922.108 process ran on assigned bx_uemnhtbv from 09:08:46.913Z to 09:08:57.743Z, then closed normally. No second browser or retry ran. The explicit hard close was 09:35Z.

WebGL version Count Newly rejected controls
1 2 Captured matrix
1 3, 4 All six per count
2 2, 3, 4 All six per count

The six controls are ignored projection, wrong selector, wrong divisor, early division, suppressed LOD and replaced captured matrix. As in the accepted runner, these controls use the normal module and nonindexed public Draw: direct GL for the first five and owning Worker GL for captured matrix. The reused positive matrix covers both normal/compat modules, direct/indexed Draw and direct/owning Worker routes on WebGL1/2. This run does not claim additional fault-route combinations.

All 31 controls reach successful public Draw followed by PIXEL ORACLE rejection, with 81 public Draws and 189 pixel observations in total. Every preceding observation within each control passes. No intentional fault is credited through a setup or runtime exception. All six matrix controls hold actual batch delivery, mutate live selector/matrix/flags through COM, replace one owned matrix, deliver the batch and reject its rendered pixels. COUNT2/3 matrix faults read 0,0,0 against 20,25,30; COUNT4 reads 0,0,0 against 60,70,30, in both WebGL versions. These are deliberate corruptions.

audit.js independently recomputes every expected RGB with the unchanged rational/mip oracle and checks tolerance two. Public Surface7 Lock arrays agree with each asserted RGB. Direct observations agree with GL readPixels and have zero GL error. Worker observations require the actual legacy WebGL endpoint, matching context version, completed draws, and zero errors, fallbacks, native software completions and fallback software completions. All public texture descriptors keep PROJECTED clear.

Reused evidence and source recovery

The exact accepted positive evidence from published bf257f136ffaa4cddb89a7335a090e983f6e4fab is reused: sixteen cells, 3,712 Draws, 10,648 observations, 616 held pairs and 72 clipped Clear samples. Its original 192,034,623-byte result SHA-256 is da6582bac3cddcde051165dbc46c28f2f3c9a6ce825d39e7509f403cec0f890f. positive-reuse.json records full lossless-result verification and matching production render/module/fixture/oracle pins. Of the previous 44 pins, only this runner and the unrelated Deus Ex inspector differ; the inspector is not executed by the fixture. No production change required rerunning positives.

The actual tested Git revision is 9eeaa3842ed998304e2bdfa74ab26f735985cd76, with relevant working changes in relevant-before.diff and exact paths/full SHA-256 in source-seal.json. The historical failed runner remains recoverable from its predecessor Git revision and recorded diff; it was not overwritten in its historical evidence. Git is the source archive. No source tree archive, repeated media or duplicate module recovery was made. Both retained modules already matched the authorized hashes, so no module transfer was needed:

Module SHA-256
Normal 129a71a01a5fbc8b07d6fa242a5a3ebc6a29f5aed9fb4b6a47d9318a764c1f52
Compat 3fde0fc42ecd0d9d5096d0b9eb83cd3086cb74aab8ac3f7b9f3fd15f5783fe6a

remote-artifacts.json and recover.log verify all five closed remote artifacts by size and full SHA-256. result.json.gz losslessly retains the 3,823,759-byte result, original SHA-256 d7c16d717a7d13aaa7671c706b09be8e02227244280394d1462ff53d837d7213.

Validation and custody

Before Chrome, separate preflight passed actual Puppeteer CJS/ESM imports, dependency lock, all 44 runtime pins, HTTP hashes/isolation, both exact modules' public COM bootstrap and production recording transport. The locked full build passed and reproduced both module hashes. The slim runtime lacks the repository cache checker/index page; the local full cache graph passes (98 page / 13 Worker scripts), and the actual runtime Worker dependencies are hash pinned. Initial attempts to invoke unavailable cache/test helpers are retained as command failures, not rendered results. Correct fixed-function shader and cache tests, shader tests, plan/module checks and the isolated emitted callback regression pass. Reintroducing the old callback scope reference in memory fails with the original ReferenceError. The global evidence checker reports existing repository errors; none names this run.

After closure, all 44 remote pins match, Chrome is absent and the fixture server is closed. Box inspection confirms healthy/hydrated RUNNING / API idle, expiry 2026-11-04T09:45:35.594Z. Shutdown ownership returns to the coordinator; the worker leaves the box running. No allocation, resume, stop, source-box access, local Chrome, subworker, successor, game, Heroes II or deployment was used.

CDP identifies SwiftShader while the context strings identify Intel/Mesa; physical hardware remains unqualified. The authorized remaining-fault work is complete. Any expansion beyond the supplied-coordinate contract is separate work.