First real kernel deliverable of Phase 1: RotationKernel trait + three
impls (Identity / Hadamard / OPQ-stub) with typed RotationError.
95/95 cognitive-shader-driver tests pass under --features serve
(+15 new D1.2 tests).

crates/cognitive-shader-driver/src/rotation_kernel.rs (~330 LOC):

  RotationKernel trait — object-safe, Send+Sync+Debug:
    apply(&self, &mut [f32]) -> Result<(), RotationError>
    dim() -> u32
    signature() -> u64              # feeds CodecParams::kernel_signature
    backend() -> &'static str       # "avx512" | "stub" (never "scalar")

  IdentityRotation { dim }
    — zero-overhead pass-through; apply() is a no-op

  HadamardRotation { dim }
    — REAL in-place Sylvester butterfly, O(N log N) add/sub,
      no allocations
    — validates dim is power-of-two (Sylvester requirement)
    — Rule C compliance: stays at Tier-3 F32x16 (add/sub, not matmul;
      AMX adds no value per plan appendix §12 C)
    — rustc + target-cpu=x86-64-v4 already emits AVX-512 add/sub
      from the straight-line loop → no JIT compilation needed

  OpqRotationStub { matrix_blob_id, dim }
    — real impl plugs into D1.1b CodecKernelEngine adapter +
      ndarray::hpc::jitson_cranelift::JitEngine + tile_dpbf16ps AMX
      matmul when amx_available()
    — apply() returns OpqMatrixNotLoaded (typed error) until the
      matrix-blob loader lands

  build(&Rotation, dim) -> Result<Box<dyn RotationKernel>> factory
    — dispatches on WireCodecParams.pre_rotation variant
    — returns typed errors on dim mismatch or non-pow2 Hadamard

Tests (15 new):
  Identity: noop + dim-mismatch error
  Hadamard:
    - orthogonality: H_4 · [1,0,0,0] == [1,1,1,1] (first column)
    - H · H = n · I (applying twice scales by n, verified at N=8)
    - norm² preservation up to n× scale (verified at N=16)
    - rejects non-pow2 dim (N=6)
  OPQ stub: returns OpqMatrixNotLoaded with blob_id preserved
  build(): identity / hadamard / hadamard-dim-mismatch / hadamard-
          non-pow2 / opq-stub
  Signatures: distinct across variants, stable for same shape,
             blob-id-sensitive for OPQ

Board hygiene (CLAUDE.md Mandatory rule):
  STATUS_BOARD.md:
    D1.2 Queued → In PR

  EPIPHANIES.md PREPEND (two entries):

    1. "Thinking styles ARE codecs over the semantic field"
       (north-star forward-looking deposit, not a work item)
       — codec infrastructure IS the template for production-grade
       thinking tissue. Mapping table documents the codec→thinking
       correspondence: CodecParams↔ThinkingStyleParams,
       kernel_signature↔style_signature, token_agreement↔
       conclusion_agreement, etc. Phase 5+ drops in
       WireThinkCalibrate + ThinkingStyleKernelCache using the
       same scaffolding. Generalisation isn't porting — it's
       recognising thinking styles as a SPECIAL CASE of the
       codec pattern.

    2. "D1.2 Hadamard is pure-Rust, not a JIT-necessary primitive"
       — narrows D1.1b scope by 30-40%. Only OPQ (matmul) needs
       Cranelift JIT emission; Identity (no-op) and Hadamard
       (butterfly) stay as plain-Rust Tier-3 F32x16 paths. Rustc's
       AVX-512 codegen under target-cpu=x86-64-v4 is already
       optimal for add/sub-structured kernels.

Rules honored:
  Rule A — in-place &mut [f32] slice, no allocations in apply()
  Rule B — ndarray::simd::* not needed for these shapes; compiler
           emits AVX-512 from straight-line loops
  Rule C — Hadamard stays at Tier 3 (add/sub, no AMX benefit);
           OPQ stub will route to Tier 1 AMX when matrix loaded
  Rule D — Rotation variants come from YAML via WireRotation (D0.1)
  Rule E — kernel signature() + backend() are object-methods per
           the Wire-surface-IS-SIMD-surface pattern
  Rule F — no serialization anywhere; in-memory f32 buffer only

https://claude.ai/code/session_01SbYsmmbPf9YQuYbHZN52Zh