D1.2 rotation primitives + taxonomy/ladder/shader-vs-engine epiphanies (95/95 tests)

[AdaWorldAPI]

Summary

Second Phase 1 kernel deliverable + three strategic epiphanies from this session's "codec IS thinking" thread.

95/95 cognitive-shader-driver --features serve tests pass (+15 new D1.2 tests).

D1.2 — Rotation primitives (5c37f0c)

crates/cognitive-shader-driver/src/rotation_kernel.rs — ~330 LOC, 15 tests.

pub trait RotationKernel: Send + Sync + Debug {
    fn apply(&self, vec: &mut [f32]) -> Result<(), RotationError>;
    fn dim(&self) -> u32;
    fn signature(&self) -> u64;     // feeds CodecParams::kernel_signature
    fn backend(&self) -> &'static str;  // never "scalar" — iron rule
}

pub struct IdentityRotation { dim }              // no-op
pub struct HadamardRotation { dim }              // REAL Sylvester butterfly, O(N log N)
pub struct OpqRotationStub { matrix_blob_id, dim } // stub pending D1.1b matrix loader

pub fn build(rotation: &Rotation, dim: u32) -> Result<Box<dyn RotationKernel>>;

Key detail — Hadamard is pure Rust, not JIT-necessary. The Sylvester butterfly is a fixed-shape add/sub pattern; rustc under target-cpu=x86-64-v4 already emits AVX-512 from the straight-line loop. Per Rule C (polyfill hierarchy), add/sub stays at Tier 3 F32x16 — AMX gives no benefit here. This narrows D1.1b Cranelift scope by 30-40%: only OPQ (matmul) needs JIT emission; Identity and Hadamard stay as plain Rust.

Tests verify:

• Identity no-op + dim-mismatch error
• Hadamard orthogonality H_4 · [1,0,0,0] == [1,1,1,1] (first column)
• Hadamard H · H = n · I at N=8 (applying twice scales by n)
• Hadamard norm² preservation up to n× at N=16
• Hadamard rejects non-pow2 (N=6)
• OPQ stub returns typed OpqMatrixNotLoaded
• build() routes correctly for all three variants
• Kernel signatures distinct across variants, stable for same shape, blob-id-sensitive for OPQ

Three forward-looking epiphanies (aad6e6a)

APPEND-ONLY on EPIPHANIES.md — not current work items, forward-looking deposits for Phase 5+.

1. Thinking styles ARE codecs over the semantic field

The codec infrastructure IS the template for production-grade thinking tissue. Mapping:

Codec (shipped D0.1–D1.2)	Thinking-style analog (future)
CodecParams	ThinkingStyleParams
kernel_signature()	style_signature()
CodecKernelCache<H>	ThinkingStyleKernelCache<H>
token agreement	conclusion agreement

Generalisation isn't porting — it's recognising thinking styles as a special case of the codec pattern.

2. Resolution ladder: 64×64 > 256×257 >> 4096×4096 > 16k (user-named)

The 5-layer stack is a resolution ladder, not a layer cake:

Size	Role	HHTL stage
64×64	p64 topology (p64_bridge::cognitive_shader::CognitiveShader)	HEEL
256×257	bgz17 palette distance (O(1) semiring.distance)	HIP
4096×4096	cross-vocab / cross-context (ndarray ScanParams JIT)	BRANCH/TWIG
16 K	Fingerprint<256> bit identity (codec decoder)	LEAF

The >> between 256×257 and 4096×4096 is the big jump (~64×) — where palette-level meets vocabulary-level. Each JIT targets its own resolution; no overlap.

p64 double-check conclusion: p64_bridge::cognitive_shader::CognitiveShader at 64×64 is architecturally clean; my D1.x codec work at 16K is at a different layer of the ladder; they compose in cognitive_shader_driver::ShaderDriver.

3. Shader vs engine: statelessness is the boundary

Cognitive shader = stateless atomic compute (eye — reports current frame, no memory).
Thinking engine = stateful orchestrator (mind — assembles frames, carries persona/qualia/world_model).

engine_bridge.rs is the seam. Codec-flexibility-as-thinking lands at the engine level, not shader level — D5/Phase 5+ drops into thinking-engine mid-layer.

Board hygiene (same commit per Mandatory rule)

• STATUS_BOARD.md: D1.2 Queued → In PR
• EPIPHANIES.md: 3 PREPEND entries (taxonomy + ladder + shader/engine)

Test Plan

• cargo test --manifest-path crates/cognitive-shader-driver/Cargo.toml --features serve --lib — 95/95 pass (+15 new)
• cargo test -p lance-graph-contract --lib — 147/147 pass (unchanged)
• cargo test --manifest-path crates/jc/Cargo.toml — 6/6 pass (JC substrate proof unchanged)
• Rules A-F honored: A (in-place slice), B (rustc compiles to ndarray::simd-compatible AVX-512), C (Hadamard correctly stays at Tier 3, OPQ stub routes to Tier 1 when wired), D (Rotation variants come from YAML via WireRotation), E (trait methods expose signature + backend), F (no serialization)

https://claude.ai/code/session_01SbYsmmbPf9YQuYbHZN52Zh