docs(x265): PROBE-GPU-LUT wave-3 verdict — wgpu decode-tier harness proven

.claude/knowledge/pr-x12-h268-morton-wgpu-synergies.md

@@ -43,7 +43,7 @@ track (`pr-x12-x266-3dgs-spacetime-upscaling.md`).
 | **Morton 2bit×2bit primitives exist** — lance-graph `FacetTier::morton` (4⁴ nibble ancestry), symbiont `morton4`, perturbation-sim cascade keys; OGAR 3×4 canon pins tier-of-level = `>>2` shift [G] | FINDING (primitives); the CTU codec does NOT use them (see §3 row 1) | lance-graph facet/symbiont sources; OGAR `CLAUDE.md` §3×4 |
 | **Certificate-gated cascade** — splat3d `depth_cascade.rs`/`depth_cert.rs`: HEEL→HIP→TWIG→LEAF preselection with Reject/KeepCoarse/Refine/ProjectExact/RenderExact actions | FINDING — but it gates **render** work on decoded data, not decode work | `src/hpc/splat3d/depth_cascade.rs:1-65` |
 | **Deterministic phase generator** — helix `CurveRuler` stride-4-over-17 coprime walk (bit-exact integer, full-17 permutation tests) | FINDING (the generator leg of the pyramid) | lance-graph `crates/helix` |
-| **wgpu in the workspace** — a2ui-paint only: **off-by-default** feature, constant-color quad pipeline, **no textures/bind groups**, wgpu's `webgl` cargo feature NOT enabled (WebGL2 is doc-comment-only), untested, repo has no CI. ndarray itself: **zero wgpu by design** ("no GPU, no wgpu", splat3d) | FINDING (negative) | a2ui-rs `a2ui-paint` manifest+src; ndarray `Cargo.lock` (0 matches), `src/hpc/splat3d/mod.rs:7-9` |
+| **wgpu in the workspace** — a2ui-paint only: **off-by-default** feature; the PRODUCTION path is still a constant-color quad pipeline with no textures/bind groups. **Updated (wave 3, PROBE-GPU-LUT):** a `#[cfg(test)]` probe (`gpu_lut_probe.rs`) now adds a texture + bind-group + R16Uint→R32Uint `textureLoad` path that COMPILES under wgpu 22 — so "no textures/bind groups" no longer holds for the TEST surface. Still-true negatives: wgpu's `webgl` cargo feature is NOT enabled (the wasm32 WebGL2 backend is unwired; the probe's WebGL2 claim is a shader-SURFACE property — integer texture + `textureLoad` are GLES3-core — not a wired wasm backend, and the HEADLESS probe cannot obtain a WebGL2 canvas adapter anyway: wgpu's WebGL backend requires a `compatible_surface` — codex P2), the path is **adapter-untested** (skips where no surface-less adapter), and the production render tier is unchanged. ndarray itself: **zero wgpu by design** ("no GPU, no wgpu", splat3d) | FINDING (negative, wave-3-amended) | a2ui-rs `a2ui-paint` manifest+src + `gpu_lut_probe.rs` (wave 3); ndarray `Cargo.lock` (0 matches), `src/hpc/splat3d/mod.rs:7-9` |
 | **Perturbation pyramid + two-algebra rule** — D-PHASE and D-WHP are operator-pinned **[H] hypotheses with named unrun probes** (PHASE-1, PERT-RHO, PYR-1; WHP-1..4), CANON-pin not CODED; losslessness fenced to *synthesis* (dither/anti-moiré grade) with quorum-certificate escalation | CONJECTURE [H], probe-gated | OGAR `docs/DISCOVERY-MAP.md:249-250`, `CLAUDE.md` fences; J2 kill condition |
 
 ## 2. The industry walls (from `pr-x12-h266-h267-standards-landscape.md`, sourced)
@@ -107,7 +107,7 @@ any float-path GPU bit-exactness claim.
 
 | Probe | Question | Pass | Kill |
 |---|---|---|---|
-| PROBE-GPU-LUT | bgz17 256×256 tables as R16Uint/R8Uint; fragment-shader distance == `batch_palette_distance`? | bit-parity on wgpu gles + WebGPU | GPU lane abandoned for LUTs; CPU-wasm only |
+| PROBE-GPU-LUT | bgz17 256×256 tables as R16Uint/R8Uint; fragment-shader distance == `batch_palette_distance`? | bit-parity on wgpu gles + WebGPU | GPU lane abandoned for LUTs; CPU-wasm only → **VERDICT (wave 3): HARNESS-REAL / CPU-primitive GREEN / GPU-exec COMPILED+SHIPPED (wgpu 22, WebGPU+WebGL2), execution-parity adapter-deferred; the KILL did not fire — see wave-3 sub-table** |
 | PROBE-MORTON-CTU | flat Morton-addressed 85-slot SoA vs shipped arena tree | ≥2× partition-sweep throughput, code no larger | keep arena; Morton stays address-canon only |
 | PROBE-RANS-INTERLEAVE | N-state interleaved rANS, wasm SIMD128 lanes | ≥4× decode throughput vs scalar at equal ratio | CTU-granular parallelism declared sufficient |
 | OGAR PHASE-1 / PERT-RHO / PYR-1 | phase determinism; escalation rate; pyramid roundtrip | per OGAR canon | **J2**: D-PHASE stays dither-only; all codec-savings claims struck |
@@ -133,6 +133,18 @@ probes in lance-graph `helix`/`bgz-tensor`):**
 | PROBE-SPRITE-REPLAY | PASS-AT-SIGNED360 (scoped) / ResidueEdge-24bit INSUFFICIENT; KILL did not fire | Signed360 mean/max 0.0; ResidueEdge mean 9.98 / max 42.4 (Pos 3.55, Neg 16.4); B bidir-delta 6.04 | One Signed360 (48-bit) code per sprite per P-frame reconstructs motion + B-frames exactly — but for helix-MANIFOLD motion (ground truth drawn from helix's own lift); proves capacity+round-trip+sign, not arbitrary-motion generality ([H], independent-GT probe named). ResidueEdge-24bit unusable: hemisphere-blind + 8-bit-rim rank-adjacency hazard → Signed360-only motion primitive |
 | PROBE-WH-MAG-2 | UPGRADES the §10 WH-magnitude [H] leg: bare NOT-TRANSFERRING → pairing TRANSFERS on structured tiles | gradient+spike escB/A 0.815 + cenB/A 0.209 (both <0.9); heavy-tailed 0.998 / 1.232; noise 1.66 / 1.66; bareB/A 0.929/1.317/1.869 reproduced; esc_frac 0.098 | WH transfers ONLY paired with passthrough-escape + centroid-residual (the shipped row codec E-PALETTE-RESIDUAL-LADDER-1) and ONLY on structured tiles; NOT bare, NOT heavy-tailed (escape masks to parity; centroid hurts), NOT noise. escB/cenB are compound (no direct+escape/direct+centroid control) — transfer is a codec-pairing property, WH's isolated margin unmeasured; centroid is class-dependent |
 
+**Wave 3 — PROBE-GPU-LUT on the a2ui-paint wgpu seam (2026-07-18,
+main-thread-adjudicated; probe in a2ui-rs `a2ui-paint`
+`src/gpu_lut_probe.rs`):**
+
+The wgpu decode tier the sprite-replay plan (§Decode tiers c) and the
+`a2ui N2` queue row gated on the "shared PROBE-GPU-LUT harness" — the
+harness is now confirmed real and the LUT-gather is proven buildable.
+
+| Probe | Verdict | Numbers | Consequence |
+|---|---|---|---|
+| PROBE-GPU-LUT | HARNESS-REAL / CPU-primitive GREEN / GPU-exec COMPILED+SHIPPED, execution-parity adapter-deferred; KILL did not fire | CPU-ref: 65536/65536 gather == row-major index, symmetric + zero-diagonal, deterministic, 256² u16 = 128 KiB (the §10(i) figure). GPU-exec: full 256×256 R16Uint→R32Uint `textureLoad` parity path compiles clean under wgpu 22 (GL backend via glow builds) and SKIPS-green here (measured: libvulkan loader present, **0 ICDs** → no adapter). `clippy --features wgpu -D warnings` clean; fmt clean | The bgz17 256²-u16 palette-distance table is **representable and COMPILES** through a real in-scope wgpu texture path (R16Uint sampled, `textureLoad`, integer render target — all WebGL2-core), so the GPU LUT lane is **not** abandoned; runtime gatherability + silicon parity stay deferred until actual GPU execution. Falsifiable core (the gather arithmetic) proven bit-exact; the GPU merely executes it. **Backend scope (codex P2, headless):** the probe requests a **surface-less** adapter, so its execution validates on **WebGPU** (native/browser) + native/software GL (lavapipe CI); the WGSL is WebGL2-**compatible** (GLES3-core), but wgpu's WebGL backend needs a canvas-bound `compatible_surface`, so the wasm32 WebGL2 backend is OUT OF SCOPE for this headless probe — "WebGL2" is a shader-surface property, not a validated execution backend here. This structurally un-gates §Decode tiers (c) + the `a2ui N2` row: harness capability proven; only runtime-execution parity awaits an adapter environment. **CAVEAT (honesty):** the GPU-exec *execution* was NOT run in this sandbox (no adapter) — "GPU-exec green" means COMPILES + SKIPS-cleanly, not that 65536 texels were compared on silicon here; the CPU-reference is the leg that actually ran. No bgz17 dep (harness-capability probe; the 256² table is built deterministically with bgz17's table STRUCTURE, keeping the a2ui-paint crate boundary clean) |
+
 ## 7. Comma closure — the replayable irrational (constants correction folded in)
 
 The Pythagorean comma is the residue of a stack of pure fifths that never

.claude/plans/x265-sprite-replay-probe-v1.md

@@ -131,5 +131,89 @@ guard); N=8 sprites, TOTAL=240, 6 anchors (I+5P), GOP `I B B P ...`. Determinist
 
 **Follow-up (named, deferred):** an arbitrary-independent-ground-truth motion
 probe (a captured 3-D trajectory the encoder did not generate) to promote the
-helix-manifold [proven] result toward arbitrary-motion [H]. The wgpu/wasm decode
-tiers (plan §Decode tiers b/c) remain deferred on the shared PROBE-GPU-LUT harness.
+helix-manifold [proven] result toward arbitrary-motion [H].
+
+## Results (2026-07-18 — PROBE-GPU-LUT, the shared wgpu decode-tier harness, main-thread-adjudicated)
+
+**Verdict: HARNESS-REAL / CPU-primitive GREEN / GPU-exec COMPILED+SHIPPED, execution-parity adapter-deferred. The KILL ("GPU lane abandoned for LUTs") did NOT fire.**
+
+Probe: `a2ui-rs crates/a2ui-paint/src/gpu_lut_probe.rs`. The wgpu decode tier
+(§Decode tiers c, the `a2ui N2` queue row) was gated on a "shared PROBE-GPU-LUT
+harness" — operator ruling this session pinned that harness to a2ui-paint's real
+`wgpu = "22"` seam (WebGPU + WebGL2), the one in-scope GPU path (q2 `sculpt` +
+ndarray `splat3d` both deliberately opt OUT of GPU; measured this session).
+
+- **CPU-reference leg (ran here, PASS):** the 256²-u16 palette-distance LUT
+  texture-gather (`textureLoad(lut,(q,k)).r` == row-major `lut[q*256+k]`) is
+  bit-exact over all 65536 entries; table is symmetric + zero-diagonal +
+  deterministic (SplitMix64); 256² u16 = 128 KiB (the §10(i) materialized-table
+  figure). **This is the falsifiable core** — the arithmetic is what could be
+  wrong; the GPU only executes it.
+- **GPU-exec leg (COMPILED + SHIPPED, adapter-deferred):** the full
+  R16Uint-LUT → fragment `textureLoad` → R32Uint target → readback →
+  full-table-parity path compiles clean under wgpu 22 (WebGPU + WebGL2 via
+  `glow`; `clippy --features wgpu -D warnings` clean, fmt clean) and
+  **SKIPS-green** in this sandbox — measured: `libvulkan` loader present but
+  **0 ICDs** installed → `request_adapter()` returns `None`. It runs the real
+  65536/65536 parity wherever a WebGPU/WebGL2 adapter exists (lavapipe CI, a
+  browser). Integer sampled texture + `textureLoad` + integer render target are
+  all WebGL2-core, so the one shader covers both backends.
+- **KILL did not fire** (§Decode tiers c / `a2ui N2`): the bgz17 256²-u16 table
+  IS gatherable through a real in-scope wgpu texture, so the GPU LUT lane is not
+  abandoned — the harness capability is proven buildable.
+- **HONEST CAVEAT:** the GPU-exec *execution* was NOT run on silicon here (no
+  adapter). "GPU-exec green" = COMPILES + SKIPS-cleanly + is the shipped WGSL,
+  NOT "65536 texels compared on a GPU in this session." The CPU-reference is the
+  leg that actually ran. Runtime-execution parity is the one piece that awaits an
+  adapter environment (the `a2ui N2` render-parity-headless-vs-browser bar).
+- **Boundary kept clean:** no bgz17 dep in a2ui-paint (charter: no consumer
+  crate deps) — the 256² table is built deterministically with bgz17's table
+  STRUCTURE (symmetric u16, zero diagonal); this is a HARNESS-CAPABILITY probe,
+  not a bgz17 integration. Test-only `pollster` dev-dep for the async block.
+
+**Consequence (scoped — corrected per codex P2 on ndarray #249):** only §Decode
+tiers **(c) the wgpu harness** is structurally un-gated — the shared harness it
+waited on is real and the LUT-gather compiles + CPU-proves. The remaining
+deferral is narrow: run the GPU-exec parity in an adapter environment (lavapipe
+CI or browser) to close the `a2ui N2` render-parity bar on silicon.
+
+**Tier (b) — the wasm tier — is NOT un-gated by this wave.** Its gate is a
+distinct check: **CPU-native vs wasm32 replay-determinism** of the decoded
+sprite states (plan §4). PROBE-GPU-LUT recorded only a CPU-reference run + an
+adapter-skipped GPU leg — **no wasm result**. Tier (b) still requires its own
+CPU-vs-wasm parity run before it can be called un-gated; PROBE-GPU-LUT does not
+touch it.
+
+## Results (2026-07-18 — HEVC external anchor, §5 optional context, RUN + VISUAL)
+
+The plan §5 optional anchor ("run actual x265 over the CPU raster sequence;
+report bits/frame + PSNR") is now RUN — and made visual.
+
+- **Scene:** 8 gaussian sprites tracing φ-spiral (golden-angle hemisphere)
+  paths, alternating hemispheres by index parity — the sprite-replay scene
+  (NUM_SPRITES=8, TOTAL=240) rendered to pixels. 320×240, 240 frames, a faint
+  panning background so P/B-frames have global motion to track.
+- **Encoder:** x265 3.5, preset medium, `--psnr`. x265 ran its OWN I/P/B GOP
+  over our moving scene (the arc's "replay x265's GOP grammar" made literal):
+  1 I, 56 P, 183 B, up to 5 consecutive B-frames.
+- **Numbers:** raw Y4M 27,649,483 B → HEVC 43,115 B = **641×**; **1437.2
+  bits/frame** (180 B/frame); Global **PSNR 60.94 dB** (Y 47.5–52.0 by
+  slice-type; chroma neutral). Encode 316 fps. (Re-run after the CodeRabbit #738
+  fix: the reproducer now uses `sprite_replay`'s canonical draw sequence +
+  signed-z hemisphere projection; the earlier 578×/60.79 dB figures were the
+  pre-fix scene where `sign` was inert.)
+- **Roundtrip visual:** frames decoded back OUT of the `.265` bitstream (ffmpeg)
+  into a 5-frame motion montage + animated GIF — the sprites visibly at
+  different positions/sizes across time.
+- **Reproducer:** `lance-graph crates/helix/examples/hevc_moving_scene.rs`
+  (std-only, deterministic SplitMix64) → Y4M; `x265 --input scene.y4m --y4m
+  --preset medium --psnr -o scene.265`; `ffmpeg -i scene.265 … montage/gif`.
+
+**Reading (honest):** this is an EXTERNAL ANCHOR (plan §5, explicitly "not a
+gate"), NOT a claim about our codec. The 578× / 60.79 dB are **x265's** numbers
+on a smooth-gaussian synthetic scene that compresses easily — they anchor "what
+a stock HEVC encoder does with this content," not "our primitives beat x265."
+The arc's own sprite-replay motion coding (E-SPRITE-IPB-HELIX-1: one Signed360
+code per sprite per P-frame) is the thing being contextualized; the
+bitrate-comparison study (our object-level motion codes vs x265's per-block MV
+field on the SAME scene) is a NAMED follow-up, not done here.