docs: PROBE-SPRITE-REPLAY spec (x265 I/P/B grammar over HHTL sprites + helix motion) + two standing corrections

New plan x265-sprite-replay-probe-v1.md (operator-directed, crawl-first
scope): moving object = HHTL-anchored Gaussian-splat sprite with a helix
direction code as the motion primitive, mapped onto the x265 GOP
grammar — I = splat set at anchor; P = ONE helix code per sprite
(object-level motion replacing per-block MV search, x265's dominant
encode cost); B = parametric interpolation along the helical path.
Explicitly NOT H.268 and NOT x265 bit-parity; CPU/wasm carries
bit-exactness (sprite states from helix codes, pinned math), wgpu is
render-grade per C5/C9. The minimal wgpu harness doubles as
PROBE-GPU-LUT's missing harness. §5 probe-queue row added with pass/
KILL bands (KILL = helix object-motion collapses into a dense per-splat
MV field).

Standing corrections folded (flagged in the prior round):
- §10(i) honesty amendment: the 192B/3-cache-line tile claim is fully
  honest only under the analytic Fisher-z canon (materialized 256^2 u16
  = 128KB, L2-resident not L1D); analytic drops table residency to 8B
  and rail reads become |delta-i8| arithmetic.
- PROBE-WH-MAG-2 deferral-condition amendment: the Skip/Merge/Delta/
  Escape mode grammar already IS the per-tile escape tier (analog of
  classify_rows_by_lfd Passthrough; shader analog = FailureTicket);
  WH-MAG-2 = WH under the mode grammar.

Blackboard entry appended. Knowledge-doc suite green (117/117).

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01K3RyLEbuNSHxxB3NTTrGki

Author: claude

.claude/blackboard.md

@@ -806,3 +806,24 @@ default is x86-64-v3 (avx2) so ndarray_simd activates on avx512 builds only.
   E-H268-PROBE-WAVE-1-RESULTS + plan h268-probe-wave-v1.md Results.
   Probes: bgz-tensor probe_wh_mag / jc sig_checksum / helix
   walk_spectrum (all suites green).
+
+## 2026-07-16 (5) — sprite amortization spec'd + two standing corrections
+
+- **PROBE-SPRITE-REPLAY spec'd** (plan `x265-sprite-replay-probe-v1.md`,
+  §5 row added): moving object = HHTL-anchored splat sprite + helix
+  motion code, mapped onto the x265 I/P/B grammar (I = splat set at
+  anchor; P = one helix code per sprite, replacing per-block MV search;
+  B = parametric interpolation along the helical path). Scope guards:
+  NOT H.268, NOT x265 bit-parity — GOP-grammar replay on our primitives;
+  CPU/wasm carries the bit-exactness claims, wgpu is render-grade (C9).
+  Amortizations: motion search → address arithmetic in the Morton
+  cascade; the minimal wgpu harness doubles as PROBE-GPU-LUT's missing
+  harness. KILL: helix object-motion collapsing back into a dense MV
+  field.
+- **§10(i) honesty amendment**: the 3-cache-line tile claim holds only
+  under the analytic Fisher-z canon (materialized 256² u16 = 128KB =
+  L2-resident); analytic drops table residency to 8B and makes rail
+  reads |Δi8| arithmetic (four tiles/lane per AVX-512 register).
+- **PROBE-WH-MAG-2 deferral weakened**: the Skip/Merge/Delta/Escape
+  mode grammar already IS the per-tile escape tier; WH-MAG-2 = WH under
+  the mode grammar, not a wait for new machinery.

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

@@ -114,6 +114,7 @@ any float-path GPU bit-exactness claim.
 | WHP-1..4 (+GPU arm) | two-algebra pyramid parity | per OGAR canon | magnitude side stays CPU |
 | Plan E bench | bits/Gaussian on Mip-NeRF 360 | ≤4 bits | R-10 re-derived; web-streaming claim withdrawn |
 | a2ui N2 | wgpu `webgl` feature + texture upload, wasm32-tested | render parity headless vs browser | GPU raster tier deferred; CPU raster only |
+| PROBE-SPRITE-REPLAY | x265 I/P/B grammar over HHTL-anchored splat sprites with helix motion codes (plan `x265-sprite-replay-probe-v1.md`; NOT H.268, NOT x265 bit-parity) | replay determinism CPU==wasm on sprite states; B-frame bidirectional consistency; helix quantization bound reported | object-level helix motion collapses back into a dense per-splat MV field — the sprite amortization dies as stated |
 
 **Run this wave — h268-probe-wave-v1 (2026-07-16, reviewer-adjudicated;
 probes live in lance-graph `bgz-tensor`/`jc`/`helix`):**
@@ -393,6 +394,16 @@ bytes actually stored are magnitudes. Same object, two consumers.
   by ruling; the C1 arena-tree corrective (§3 row 1: the shipped `ctu.rs`
   is a pointer arena, not Morton-flat) applies to the **codec CTU**, not
   to the L4 substrate — the two do not contradict each other.
+  **Honesty amendment (2026-07-16, post Fisher-z canon):** the
+  3-cache-line claim is fully honest only under the **analytic**
+  cosine-replacement (E-FISHERZ-CANONICAL-COSINE-REPLACEMENT-1) — a
+  materialized 256² u16 distance table is 128 KB (L2-resident on typical
+  parts, NOT L1D), so under the materialized flavour every rail read was
+  silently L2-bound and the working set was never really 3 lines. The
+  analytic codec drops table residency to the 8-byte `FamilyGamma`,
+  making the per-tile claim true as stated; similarity reads become
+  arithmetic-shaped |Δi8| (64 lanes/AVX-512 register = four tiles' worth
+  of one lane per register).
 
 **The four-role loop:** **φ PLACES → walk QUANTIZES → γ CORRECTS → F
 DECIDES.** λ-RDO (rate-distortion optimization, the codec's tile-local
@@ -440,6 +451,13 @@ fourth-mode/anchor discussion (§8):
   row-level win does not transfer to per-tile granularity, and the
   shipped row codec is untouched. Follow-up PROBE-WH-MAG-2 (WH + escape
   tier + centroid residual) deferred to a per-tile-codec consumer.
+  **Deferral-condition amendment (2026-07-16):** the codec's 2-bit mode
+  grammar (Skip/Merge/Delta/**Escape**, `mode.rs`) already IS a per-tile
+  escape tier — the tile-granular analog of `classify_rows_by_lfd`'s
+  Passthrough (and the shader-side analog is F-dispatch's
+  FailureTicket). PROBE-WH-MAG-2's natural home is therefore WH
+  re-evaluated UNDER the mode grammar (heavy-tailed tiles routed to
+  Escape before the cascade), not a wait for new machinery.
 - **[S] Signature as the replayable-trajectory checksum.** The x264
   contrast in §7 asked for a replayable, checksummable stream; the
   Hambly–Lyons signature is the canonical path digest with a uniqueness

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

@@ -0,0 +1,95 @@
+# PROBE-SPRITE-REPLAY v1 — x265 I/P/B grammar over HHTL sprites with helix motion
+
+> Date: 2026-07-16. Status: SPEC'D (operator-directed; execution wave is
+> the next plateau after TD-BGZ-TENSOR-PRE-LANE-REVIEW lands).
+> Scope guard, operator's words: the initial test is NOT reaching for the
+> stars (H.268) — it is **a simple replay of x265's GOP grammar** using
+> our own primitives, on our own hardware tiers.
+
+## The thesis (the new amortization)
+
+**A moving object = an HHTL-addressed spatial sprite** — a Gaussian-splat
+set anchored at an HHTL address — **with a helix direction code as its
+motion primitive**, mapping directly onto x265's basic frame operations:
+
+| x265 op | Sprite equivalent | Primitive |
+|---|---|---|
+| **I-frame** | the sprite's full splat set at its anchor address | splat3d EWA set + HHTL anchor (HEEL\|HIP\|TWIG) |
+| **P-frame** | ONE helix motion code per sprite (object-level), replacing the per-block MV field | helix `ResidueEdge` (24-bit hemisphere) or `Signed360` (48-bit signed full-sphere) |
+| **B-frame** | parametric interpolation along the helical path between two anchors | evaluate the sprite at t ∈ (0,1); bidirectional weights |
+
+The amortization stack:
+- **Motion search dies.** x265's dominant encoder cost (per-block MV
+  search) becomes address arithmetic: the sprite's anchor moves; its
+  splats ride along. The nested stacked-inverse-pyramid 4×4 ergonomics
+  re-rasterize the sprite footprint through the Morton cascade at
+  whatever LOD the certificate demands (depth_cascade actions).
+- **One substrate, both consumers** (per E-H268-REPLAYABLE-TILE-1): the
+  same tile cascade serves codec rasterization and shader dispatch.
+- **The wgpu harness is shared.** The minimal render harness this probe
+  needs is exactly the harness PROBE-GPU-LUT has been gated on — build
+  once, both probes consume it.
+- **Fisher-z canon applies** (E-FISHERZ-CANONICAL-COSINE-REPLACEMENT-1):
+  similarity/direction-adjacent reads carry as normalized i8; helix runs
+  the 2z rung of the same analytic family (`batch_fisher_z` exists).
+
+## Scope guards (graded, non-negotiable)
+
+- **NOT H.268.** No scene-codec claims, no beyond-VVC claims.
+- **NOT bitstream/byte parity with x265.** "Replay of x265" = replaying
+  the **I/P/B operational grammar** with our motion primitive; x265
+  itself is at most an optional external reference point (bitrate/PSNR
+  context on the rasterized sequence), never a parity gate.
+- **GPU is render-grade only** (C5/C9 discipline): the bit-exactness
+  claims live on the CPU/wasm integer/pinned-math path (sprite STATES:
+  helix codes → positions); raster output compares to tolerance, GPU
+  raster is a visual tier.
+- All claims [S/H] until the probe runs; this plan is the spec, not a
+  result.
+
+## Test spec (minimal, deterministic)
+
+1. **Scene:** N=8 sprites × K=64 Gaussians each (seeded), moving on
+   ground-truth helical paths (the helix codes ARE the ground truth —
+   encode direction as `ResidueEdge` AND `Signed360`, measure both
+   widths' quantization error).
+2. **Encode:** GOP = I B B P B B P … (classic pattern): I = splat dump +
+   anchors; P = per-sprite helix delta codes; B = no stored motion —
+   derived by parametric interpolation between surrounding anchors.
+3. **Decode tiers:** (a) CPU native (ndarray `splat3d` EWA rasterizer);
+   (b) wasm (same code — the shipped parity-CI pattern); (c) wgpu quad/
+   splat raster (a2ui-paint tier; N2 gate applies).
+4. **Pass criteria:**
+   - **Replay determinism:** decoded sprite states (positions from helix
+     codes, pinned unfused math) bit-identical CPU native vs wasm.
+   - **B-consistency:** a B frame decoded forward-from-I and
+     backward-from-P agrees with the parametric midpoint to a stated
+     tolerance (the bidirectional check).
+   - **Motion fidelity:** helix-coded direction reproduces ground-truth
+     paths within the register's quantization bound (report 24-bit vs
+     48-bit error curves).
+   - **KILL:** if object-level helix motion cannot express the test
+     paths without per-splat residual fields (i.e. the "one code per
+     sprite" claim collapses back into a dense MV field), the sprite
+     amortization dies as stated and the finding is the honest record.
+5. **Optional context (not a gate):** run actual x265 over the CPU
+   raster PNG sequence; report bits/frame + PSNR as an external anchor.
+
+## Standing corrections folded from the probe wave (same doc pass)
+
+- §10(i) cache-honesty: the 192 B/3-cache-line tile claim is fully
+  honest only under the analytic Fisher-z canon (a materialized 256² u16
+  table is 128 KB — L2-resident, not L1D); the analytic path drops table
+  residency to 8 B.
+- PROBE-WH-MAG-2's deferral condition is weaker than recorded: the
+  codec's 2-bit mode grammar (Skip/Merge/Delta/**Escape**) already IS
+  the per-tile escape tier the probe lacked; re-running WH under the
+  mode grammar is the natural PROBE-WH-MAG-2 home.
+
+## Execution model
+
+Same wave pattern: drafters (grindwork: scene generator, GOP encoder,
+CPU/wasm decode, harness), filigree adjudication vs the pass/KILL bands,
+central gates, PR, autonomous merge (standing authority). The wgpu tier
+may land as a second commit gated on the shared harness; if it slips,
+the CPU/wasm probe stands alone.