Filtered P8 index identity — 2026-10-20

The all-nonkey filtered P8 false rejection is repaired. The span retains its existing FALSE-mode rejection expression, then cancels that rejection only when an ordinary opaque P8 footprint contains no native key indices. The membership helper uses the sampler's float32 half-texel footprint, nonfinite canonicalization, addressing and original byte indices. All four fetched taps participate in this absence check, including zero-weight taps. It does not introduce an any-key-tap rejection rule.

Point identity, explicit TRUE replacement/bypass, alpha palettes, GL border taps and 16-bit paths retain their existing behavior. Alpha palettes and GL border addressing remain outside this bounded identity path. Mixed footprints retain the old result exactly; their tests are characterization, not native FALSE-mode conformance.

Source reconciliation and reproducer

Initial main: 63d22caf879ca77d891d284f7926416a210136c2. Fresh main at final review: f5b447ebbc2b047451926053a33b367d578b4a36; its intervening five files are contract reports/checks only, with identical renderer/compiler blobs. Read report/check 1914ce3492099868c4b4f9a2180b35fbfa093543.

The prior report supplied a red key_rgb to its interpreted expression. Executing the actual span exposes a narrower, exact reproducer: key index 0, palette entries 0 and 1 both black, texture indices [1,1,1,1], linear sampling at (0.5,0.5). The original-index membership oracle is false for every fetched tap. Baseline leaves framebuffer 0x12345678 and depth 1; the fix writes 0xff000000 and depth 0.25.

Production's existing d3dim_decode_surface_pixel does not look up a P8 key in the palette: its fallback ORs the raw key with 0xff000000. Consequently an all-key red footprint with key 0 survives on both baseline and patched source. This separate existing all-key decoder defect is retained as an explicit reproducer, not reported as a conformance pass or corrected by guessing a broader key/filter policy. All-key black rejection is verified. Changing the decoded key RGB also changes mixed-footprint outcomes and is outside this narrowly scoped all-nonkey repair.

The regression also verifies nonzero key 1 with duplicate RGB 0x000001, and distinct nonkey colors 0x000000/0x000002 filtering to 0x000001. No palette-RGB-to-index reverse mapping is used.

Executed checks

From the task checkout:

node --check docs/re-notes/d3d-filtered-index-identity-20261020.js
node docs/re-notes/d3d-filtered-index-identity-20261020.js /absolute/evidence/directory
git diff --check

The JS check extracts and compiles the complete actual production span and sampler closure with the vendored WATX production compiler, then executes the WebAssembly in Node/V8. It verifies 134 cases: duplicate/nonduplicate P8, point/linear, key disabled and absent, TRUE bypass, all-key/all-nonkey, all 16 native membership masks, mixed/zero-weight footprints, wrap/mirror/clamp, nonfinite coordinates, and RGB565/XRGB1555/ARGB1555 fast16 versus forced general sampling. Assertions check framebuffer bytes and depth writes. Original-index membership is checked separately on all masks. Baseline and guard-removal mutation both reproduce the duplicate-black rejection.

Surface format/palette side-table access uses explicit fixture adapters; guest-to-Wasm is identity for the test depth buffer. Execute-buffer extent bookkeeping is a no-op. Disabled clipping, second texture, fog, blending, stencil, dithering and border helper paths trap if invoked. Production color-stage combination, sample decoding, key processing, alpha predicate, depth comparisons and span framebuffer stores execute unchanged. The forced-general variant changes only the fast16 branch condition. Exact retained functions and adapters are listed in tests.json.

SHA256 identities:

Evidence: scratch/runs/20261020-d3d-filtered-index-identity-oct20/. It retains four WAT/Wasm closures, source/compiler identities, numeric observations, initial harness errors, patch, preflight and final receipt. The initial harness required missing dormant stencil adapters and correction to the compiler's production input mode; these were fixture errors, not renderer failures.

Remaining obligations and scope

No browser, GPU shader execution, full-module build, full VM or native runtime was run. GPU upload/cache/shader agreement and full-VM opaque P8 draw/depth behavior still need separately authorized validation. The known nonblack all-key decode defect needs its own bounded assignment. State144 FALSE mixed/zero-weight semantics, native initial value, Device3 compatibility and alpha-palette/key precedence remain unresolved exactly as in the contract reports. The recorded resume403 was not retried or bypassed.

Operational deviation: sparse setup initially failed because its metadata info directory did not exist, but the following checkout still executed, temporarily materializing 245 MiB. This violated the 8 MiB cap and patterns-before-checkout requirement. Immediately corrected to exact sparse paths before production edits; observed free space at the oversized checkout was 2,680,438,784 bytes, above floor 2,210,398,208 plus 32 MiB reserve and 24 MiB concurrent caps. The automatically created own upstream stanza was removed; shared HEAD/index/config hashes match preflight. No unrelated WIP, laptop HeroesII/WarcraftIII or hold was edited. Final receipt accounts for task checkout, worktree metadata and retained evidence.