Final Phase 1 scaffold deliverable. Hydration/calibration path, NOT
cascade inference path (per cognitive-shader-architecture.md line 582:
the cascade uses p64_bridge::CognitiveShader::cascade + 8 predicate
planes × bgz17 O(1) distance, no per-inference codec work).

crates/cognitive-shader-driver/src/decode_kernel.rs (~280 LOC):

  DecodeKernel trait — object-safe, Send+Sync+Debug:
    decode(&self, &[u8]) -> Result<Vec<f32>, DecodeError>
    encode(&self, &[f32]) -> Result<Vec<u8>, DecodeError>
    bytes_per_row() -> u32
    dim() -> u32
    signature() -> u64          # JIT cache key
    backend() -> &'static str   # never "scalar" on SoA

  StubDecodeKernel { dim, tag } — byte-exact f32 ↔ u8 round-trip via
    native-endian reinterpret. No quantization, no compression; exists
    so composition plumbing can be tested without a trained palette.
    Backend = "stub". Signature hashes "stub_decode" + dim + tag.

  ResidualComposer { base: Box<dyn DecodeKernel>, residual: Box<dyn DecodeKernel> }
    Two-stage residual composition:
      encode(v) = [base.encode(v); residual.encode(v - base.decode(base.encode(v)))]
      decode(enc) = base.decode(enc[..base_b]) + residual.decode(enc[base_b..])
    Nests recursively — residual slot can itself be a ResidualComposer
    (depth > 1). Rejects mismatched dims at construction.
    Backend = "stub" if either stage is stub, else base's backend
    (weakest-link reporting for latency-critical stages).

  DecodeError { SizeMismatch { expected, actual }, Stage { stage, detail } }

Tests (9 new, all under --features serve):
  - stub_round_trip_is_exact
  - stub_rejects_wrong_input_size
  - residual_compose_round_trip_is_exact_when_both_stubs
    (both stubs = byte-exact; residual all zero; output == input)
  - residual_compose_mismatched_dims_rejected
  - residual_compose_bytes_per_row_sums_stages
  - residual_compose_nested_depth_two_round_trip
    (ResidualComposer whose residual IS another ResidualComposer —
    depth=2 encodes 3 stages, still byte-exact when all stubs)
  - signatures_distinguish_composer_from_stages
  - signature_depends_on_stage_order
    (base+residual vs residual+base produce different signatures)
  - composer_backend_reports_stub_when_any_stage_is_stub

Scope clarification (per orientation loaded this session from
cognitive-shader-architecture.md + ripple-dto-contracts.md):
  - D1.x codec kernels = hydration/calibration path
  - Cascade inference path = p64_bridge::CognitiveShader at L2
  - Real kernels replace StubDecodeKernel once D1.1b lands the
    ndarray::hpc::jitson_cranelift::JitEngine adapter

Board hygiene (CLAUDE.md Mandatory rule):
  STATUS_BOARD.md D1.3 Queued → In PR

Rules honored:
  Rule A — in-place &mut operations via Vec; no manual index math
  Rule B — no std::arch / no hpc::simd_avxNNN reach
  Rule C — n/a at the composition layer (real kernel backend selection
           defers to D1.1b per-stage)
  Rule D — codec params come from CodecParams via Wire DTOs (D0.1-D0.3)
  Rule E — trait methods expose signature + backend + bytes_per_row
  Rule F — no serialization between stages; Vec<f32>/Vec<u8> owned

https://claude.ai/code/session_01SbYsmmbPf9YQuYbHZN52Zh