All consumer code uses crate::simd only. Zero raw intrinsics.
LazyLock dispatch table selects AVX-512 vs AVX2 at startup.

cam_pq.rs — squared_l2():
  - Called 1,536× per CAM-PQ query (6 subspaces × 256 centroids)
  - Was: scalar iter().zip().map().sum()
  - Now: F32x16 for 16D subvectors (one SIMD lane = one subspace dimension)
  - Fast path: n==16 → single load-subtract-multiply-reduce
  - Medium path: n>=16 → chunked F32x16 with mul_add + scalar remainder
  - Estimated 16× speedup on hot path

deepnsm.rs — nsm_decompose() normalization:
  - Was: scalar iter().sum() + scalar /= loop
  - Now: F32x16 accumulation (4×16=64 elements) + scalar remainder (10)
  - Normalize via F32x16 * splat(1/sum) + scalar tail

deepnsm.rs — nsm_to_fingerprint() XOR:
  - Was: scalar for j in 0..1250 { result[j] ^= pattern[j] }
  - Now: U8x64 XOR (19×64=1216 bytes) + scalar remainder (34 bytes)
  - 64 bytes per SIMD operation vs 1 byte scalar

deepnsm.rs — nsm_similarity() cosine:
  - Was: scalar 3-accumulator loop over 74 elements
  - Now: F32x16 with mul_add for dot/mag_a/mag_b (4×16=64) + scalar tail (10)
  - Three reductions in one pass

23 deepnsm tests + 7 dispatch tests passing. Zero regressions.

https://claude.ai/code/session_01Y69Vnw751w75iVSBRws7o7

Document the category-padded SoA layout (16 categories × 16 slots = 256
F32x16 lanes) as a future optimization concept in deepnsm.rs.

Verified no overlap with existing patterns:
  - blasgraph CSR/CSC: graph adjacency matrix, not semantic vectors
  - SPO semiring: cost algebra, not vector layout
  - neighborhood CLAM: search scope, not decomposition format
  - aabb/spatial_hash SoA: spatial coords (x,y,z), not semantic categories
  - dn_tree SoA: HV summary layout, not NSM category decomposition

The concept is clean for future implementation.

https://claude.ai/code/session_01Y69Vnw751w75iVSBRws7o7

…ling spec

Two architectural concepts saved for dedicated implementation sessions:

1. Psychometric validation for DeepNSM measurement instrument:
   - Cronbach's α across 128 projections (2³ SPO × 2⁴ HHTL)
   - Split-half reliability: Strategy A vs Strategy B distance
   - IRT item parameters: per-word difficulty + discrimination
   - Factor analysis: do 74 primes factor into 16 NsmCategory?
   - Construct/convergent/discriminant validity across codec chain
   - Polysemy detection via α drop across projections
   - P-values with 128 independent measurements per pair

2. Vertical HHTL bundling (studio mixing analogy):
   - Leaves → bundle → Twigs → bundle → Branches → bundle → Hip
   - Each level = majority vote denoising (background noise removal)
   - Unbind bottom-up to verify reconstruction (information loss audit)
   - Combined SPO × HHTL = 128-way factorial decomposition
   - Cascade as psychometric filter: discrimination, factor analysis,
     composite reliability, SEM, residual analysis

Key insight: NARS confidence IS measurement reliability (formalized).
Every similarity judgment gets a confidence interval backed by
128 independent projection measurements.

https://claude.ai/code/session_01Y69Vnw751w75iVSBRws7o7