Opaque P8 all-key rejection — 2026-10-20

The compiled software span now rejects a filtered opaque P8 footprint when all four fetched native indices equal the surface key, independently of palette RGB. It preserves framebuffer and depth instead of painting the key. This is a bounded endpoint correction, not a resolution of mixed filtering.

Base: reviewed 3613e8afe93525fbf37cbeb2ebd4eae261a82c12, independently accepted by a8ece350e92f958d8f0c0fff407f5b2fa4f634e0. Fresh main before source edits was ad171b162eec9419bf97ecaea179879264c1ec19. Its renderer differs from the base only by the predecessor's nonkey identity repair; GL raster and all five compiler files are identical. This task stays on its own branch. No main integration was performed.

Diagnosis and bounded contract

d3dim_decode_surface_pixel delegates to the format decoder. For P8 that decoder falls through to raw_key | 0xff000000; it does not look up the palette. The filtered FALSE rejection compares sampled RGB against that raw value. With key 0 and red palette entry 0, four keyed taps therefore write 0xffff0000 and depth 0.25. The original predecessor evidence reproduced this on both its baseline and its fixed source.

The primary-contract report docs/re-notes/d3d-state144-contract-20261020.md at main, also retained under scratch/runs/20261020-d3d-state144-contract-oct20/REPORT.md, establishes the narrow all-fetched-key rejection and all-fetched-nonkey draw endpoints for ordinary opaque P8. Its native FALSE mixed/zero-weight algorithm, initial state144, Device3 compatibility and alpha-palette precedence remain unresolved. No new native observation or external contract search is claimed here.

Changing the shared decoder to look up P8 palette entries would change the RGB comparator for mixed footprints. Instead, the existing footprint helper's identical coordinate/address arithmetic now produces a four-bit native-key mask. The existing membership helper wraps mask != 0, preserving the reviewed nonkey guard. An added span guard rejects only mask 15, only when the existing opaque P8 identity mode applies, filtering is enabled, TRUE replacement is disabled and the fragment is still eligible. All four fetched taps count, including zero-weight taps. A mixed mask, even when only its nonkey taps have zero weight, keeps the previous behavior.

Point keying, alpha-bearing palettes, GL border routing, disabled/absent surface key, TRUE replacement and 16-bit paths keep their prior logic. The decoder and every existing production function except the span and membership helper are asserted unchanged by the regression. Alpha and depth tests still occur after key rejection. The mask contains bits, not logical AND operands; the new guard's logical operands are normalized predicates.

Executed regression

From the task repository root, with an existing output directory:

node --check docs/re-notes/d3d-allkey-decode-20261020.js
node docs/re-notes/d3d-allkey-decode-20261020.js /absolute/output > /absolute/output/tests.json
git diff --check

The dedicated JS regression derives its closure harness from the predecessor, without changing that predecessor or its evidence. It compiles actual source functions through the vendored WATX production compiler and executes them in Node v24.18.1. It does not substitute a JS key-decision model.

• 1,564 span cases pass, checking framebuffer, depth and sample. Coverage includes black and nonblack keys, duplicate RGB, native keys 0/1/7/255, all-key/nonkey/all 14 mixed masks, distinct mixed colors, exact-zero and small-positive weights, wrap/mirror/clamp, nonfinite coordinates, point/TRUE/disabled-key paths, palette-alpha exclusions, actual GL border sampling, and three 16-bit texture formats compared against the forced-general path.
• The red repro retains framebuffer 0x12345678 and depth 1 with the fix. Both the reviewed baseline and a mutation removing only the added all-key rejection write 0xffff0000 and depth 0.25. The same control exposes nonblack nonzero-key defects, except the intentional equality case: raw key 255 and RGB blue 255 already reject in baseline.
• 1,533 removed-fix differential cases exactly reproduce the reviewed baseline, proving the helper refactor itself does not change span output. The earlier nonkey guard-removal mutation still exposes duplicate-black nonkey rejection; the actual fix continues to draw those nonkey indices.
• 15,552 independently traced sampler footprints / 746,496 mask checks pass. Only the sampler fetch boundary is instrumented to log addresses; its production coordinate arithmetic is unchanged. The matrix uses 3x5, 4x2 and 1x1 textures with padded rows, all 16 address-mode pairs, 18 coordinates per axis including float32 rounding and saturation boundaries, and single-key/single-nonkey placements for keys 0/7/255. The returned mask and original boolean membership both agree with actual fetched addresses.
• Additional controls preserve alpha NEVER rejection and the existing TRUE-plus-alpha-palette trap on fixed, baseline and removed-fix modules.

Format/palette metadata access uses fixture adapters; guest depth translation is identity and extent bookkeeping is a no-op. Disabled clipping, fog, blending, second texture, stencil and other unrelated paths trap. Real sampling, decoding, stage combination, alpha predicates, depth comparison, framebuffer/depth stores and GL border sampling execute. The render target is 32-bit; the 16-bit coverage refers to textures. The retained functions, adapters, source/compiler hashes and numeric observations are in tests.json.

The baseline closure hash differs from the predecessor's harness because this dedicated regression retains real GL border helpers instead of trap stubs. The production baseline core remains the exact reviewed object.

Recovery, limits and stopping point

Evidence: scratch/runs/20261020-d3d-allkey-decode-oct20/, including tests.json, six WAT/Wasm closures, setup/preflight, recovery patch, receipt and hashes. The first control expectation incorrectly assumed every nonblack RGB differed from a raw key; its failure is retained as initial-control-expectation.stderr. It was a fixture assertion correction; no production change followed that failure.

The fresh independent repository uses read-only object alternates and exact sparse patterns installed and verified before checkout. Admission includes the 2,210,398,208-byte floor, 32 MiB reserve and 24 MiB concurrent caps. Own checkout, metadata and evidence remain under 8 MiB. Only this report, dedicated JS and the core source are committed; original evidence is retained in place. Recreate ignored compiled artifacts with the command above.

No full build, full VM, browser, GPU, native driver, benchmark, Boat access, worker launch, public deployment or held diagnostic was run. Real surface metadata/allocation, GPU upload/cache/shader agreement and end-to-end draw qualification remain untested. FALSE mixed/zero-weight policy, initial state144, Device3 and alpha/key contracts remain open. This task stops with the tested branch for independent review, without main integration.