feat: phyllotactic-manifold + p64 crates (additive, 44 tests)

[AdaWorldAPI]

feat: phyllotactic-manifold + p64 — 7+1 φ/γ spiral + 64×64 BNN palette

Two new standalone crates. Additive only — zero changes to existing files.

phyllotactic-manifold (876 LOC, 14 tests)

7+1 phyllotactic SIMD manifold with Euler-gamma singularity correction.

• 4 encoding strategies: Flat8 / Spiral8 / Spiral8Gamma / SevenPlusOne
• Rust 1.94: std::f64::consts::GOLDEN_RATIO + EULER_GAMMA
• 7 lanes on φ-spiral (prime = zero aliasing) + 1 contradiction channel
• Gap CV improvement: 49.5% (√n → √(n+γ))
• Dead Zone: contradiction displacement = 0.000 exact
• NARS truth extraction: <f, c> from resonance + contradiction

p64 (1817 LOC, 23 tests)

Palette64: 64×64 BNN attention matrix from 8 phyllotactic HEEL planes.

• 512 bytes = 8 cache lines, #[repr(align(64))]
• HeelPlanes → expand() via golden shift 39 (gcd(39,64)=1)
• BNN attention: Q AND K → POPCNT → Gamma threshold
• MoE fanout: 8 expert gates, soft/hard combination
• Denoising: convergence in 2 steps to fixed point

sparse256 module:

• 256×256 HHTL via CLAM triangle inequality pruning
• BSR SpMV: 32× speedup vs dense (3.12% FLOP ratio)
• HHTL cascade: 4 of 256 computed (64× pruning)

Palette3D (4KB L1-resident):

• 8 layers × 64×64 = predicate-typed attention
• 6 ThinkingStyles: ANALYTICAL / CREATIVE / FOCUSED / INTEGRATIVE / DIVERGENT / META
• 4 CombineModes: Union / Intersection / Majority / Weighted
• 4 ContraModes: Suppress / Ignore / Invert / Tension
• 3 deduction rules: CAUSES×ENABLES / SUPPORTS² / REFINES-chain
• Dynamic masking: palette GROWS during inference

Connection to existing stack

CausalEdge64 u64 layout → palette block addressing (S→row/4, O→col/4)
Pearl causal mask 3-bit  → 8 predicate layer selection
NARS InferenceType       → deduction rule mapping
HdrSemiring              → CombineMode + ContraMode
ThinkingStyle 12×3=36    → LazyLock<[ThinkingStyle; 12]>
DeepNSM codebook 256     → Palette64 interaction matrix
SimdCaps LazyLock        → avx512vpopcntdq dispatch
F64x8 / U64x8           → manifold expansion + batch attend

Bridge code (p64_bridge.rs, 695 LOC) ready for integration into ndarray --features p64 or as lance-graph consumer crate.

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 750ca413c5

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

[chatgpt-codex-connector]

Make denoise_step honor the noise_level parameter

denoise_step passes noise_level into attend, but attend uses gamma only to build the fires bitmask and not to choose best_idx, so the returned row is identical for all noise_level values. This means the denoising schedule in denoise has no effect on the actual state updates, which breaks the intended coarse-to-fine behavior for every nontrivial denoising call.

Useful? React with 👍 / 👎.

[chatgpt-codex-connector]

Track sparse256 active_blocks at block granularity

from_clam_leaves increments interactions (and pruned) inside the inner 4×4 leaf-pair loop, but later publishes these as active_blocks/pruned_blocks. That counts leaf pairs rather than blocks, so reported block statistics can exceed total_blocks and misstate sparsity, which can mislead downstream tuning or pruning decisions that rely on these metrics.

Useful? React with 👍 / 👎.