From cbbaf0b87f618b6a94010b97088da363e2843544 Mon Sep 17 00:00:00 2001 From: Claude Date: Thu, 16 Jul 2026 10:31:31 +0000 Subject: [PATCH] docs: PROBE-SPRITE-REPLAY spec (x265 I/P/B grammar over HHTL sprites + helix motion) + two standing corrections MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit 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 Claude-Session: https://claude.ai/code/session_01K3RyLEbuNSHxxB3NTTrGki --- .claude/blackboard.md | 21 ++++ .../pr-x12-h268-morton-wgpu-synergies.md | 18 ++++ .claude/plans/x265-sprite-replay-probe-v1.md | 95 +++++++++++++++++++ 3 files changed, 134 insertions(+) create mode 100644 .claude/plans/x265-sprite-replay-probe-v1.md diff --git a/.claude/blackboard.md b/.claude/blackboard.md index 63590010..1c156b2b 100644 --- a/.claude/blackboard.md +++ b/.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. diff --git a/.claude/knowledge/pr-x12-h268-morton-wgpu-synergies.md b/.claude/knowledge/pr-x12-h268-morton-wgpu-synergies.md index fe203c1a..77009e19 100644 --- a/.claude/knowledge/pr-x12-h268-morton-wgpu-synergies.md +++ b/.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 diff --git a/.claude/plans/x265-sprite-replay-probe-v1.md b/.claude/plans/x265-sprite-replay-probe-v1.md new file mode 100644 index 00000000..7f24ce5a --- /dev/null +++ b/.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.