Independent all-key review, 2026-10-20

ACCEPT 0571e3d08272e12bef9cd7bf43d6bcc763c6e649 atop accepted 3613e8afe93525fbf37cbeb2ebd4eae261a82c12 for the bounded opaque-P8 all-four-fetched-native-key rejection. No production fixes or main integration.

The exact three-file diff adds one four-bit helper, converts the old boolean membership helper to mask != 0, and rejects mask 15 in the span before alpha/depth processing. Each mask bit uses precisely the previous helper's address calculation and unsigned byte comparison. The wrapper preserves the old boolean result for every mask. The guard requires an eligible fragment, effective opaque P8 identity mode, nonzero filtering and disabled TRUE replacement. An already rejected fragment stays rejected. The earlier all-nonkey repair remains effective. Shared raw-key decoding is byte-for-byte unchanged, as are all other existing functions except the span and membership helper. GL raster source is unchanged.

Thus mixed masks retain the existing comparison, including zero-weight taps. Point sampling, no effective key, TRUE replacement, alpha-bearing palettes, GL border exclusion and 16-bit textures cannot enter the new rejection path. This establishes preservation, not correctness of unknown native policies. The accepted state144 contract report at the candidate commit explicitly limits its FALSE endpoint conclusion to basic all-key transparency; it does not establish native mixed/zero-weight/default or Device3 behavior. This review makes no new native/API-conformance claim.

Executed evidence:

• All 25 seals in the original SHA256SUMS.json verified in place before testing; original evidence was never modified.
• Actual production-compiler closure rerun: 1,564 span cases passed; output JSON is byte-identical to original evidence. Baseline and all-key-removed controls write red and depth 0.25; fixed preserves destination and depth. All 1,533 removed-fix differential cases match baseline. The previous nonkey guard-removal control still fails the intended duplicate-black case.
• 15,552 actual sampler footprints and 746,496 mask checks passed again. Fetch-boundary instrumentation retains production coordinate arithmetic.
• Independent helper executed 9,216 additional actual span cases against baseline and removed-fix modules. It challenges a concrete coverage gap: original all-key span cases mostly use uniform 2x2 textures, whereas mask tracing alone does not test span wiring on other dimensions. New 3x5, 1x4, and 5x1 textures contain nonkey texels outside an all-key local footprint. All 3,072 such cases reject, while baseline and removed-fix write. Changing one fetched texel yields mixed/repeated-address cases; 2,248 mixed and 3,896 all-nonkey cases preserve baseline output. All 16 address mode pairs and eight independent coordinates per axis are exercised. Expected footprint membership comes from actual sampler fetch addresses, not the new key-mask helper. No counterexample found.

SHA-256 identities:

Evidence directory: scratch/runs/20261020-d3d-allkey-independent-review-oct20. tests.json records every compiler/source hash and numeric observation; independent.json records independent case counts; original-seals.json records original manifest identity and all verified hashes. Final receipt and seal manifest record review commit, verified remote and resource accounting.

Recovery from the candidate plus this report/helper commit, in a repository containing the five compiler files and the two production source fragments:

node docs/re-notes/d3d-allkey-decode-20261020.js /existing/output > /existing/output/tests.json
node docs/re-notes/d3d-allkey-independent-review-20261020.js /existing/output > /existing/output/independent.json

Fresh isolated Git repository uses shared objects read-only through alternates. Exact nine-file sparse scope was planned from the diff, installed and verified before checkout, then checked against materialized index paths. Only this report/helper are committed. Shared HEAD/index/config and original evidence are rechecked at completion. No production/main edits, full build, full VM, GPU, browser, benchmark, runtime access, Boat/SSH commands, workers, held diagnostic, deployment or laptop game work occurred.

Limits: fixture metadata accessors and identity guest translation; render target is 32-bit (16-bit coverage concerns textures). Disabled unrelated paths trap, extent bookkeeping is a no-op. Native mixed/zero-weight/default, Device3, alpha-key contracts, real metadata allocation, GPU/cache/shader and end-to-end draw qualification remain open. Integration gates are outside this review; the board separately records the predecessor's integration blocked on the existing Jedi raw-literal gate.