diff --git a/.claude/knowledge/cognitive-shader-foundation.md b/.claude/knowledge/cognitive-shader-foundation.md
new file mode 100644
index 00000000..76426cd5
--- /dev/null
+++ b/.claude/knowledge/cognitive-shader-foundation.md
@@ -0,0 +1,137 @@
+# Cognitive Shader Foundation — ndarray's Role in the 7-Layer Stack
+
+> READ BY: all ndarray agents (arm-neon-specialist, cascade-architect,
+> cognitive-architect, l3-strategist, migration-tracker, product-engineer,
+> savant-architect, sentinel-qa, truth-architect, vector-synthesis)
+>
+> Parallel doc in lance-graph: `.claude/knowledge/cognitive-shader-architecture.md`
+
+## ndarray's Role: Layer 0 + parts of Layer 1
+
+ndarray is the HARDWARE FOUNDATION of the cognitive shader stack.
+It provides the primitives that Layer 1 (BindSpace columns) and higher
+layers build on. ndarray never depends on lance-graph or any cognitive
+crate — the dependency is one-way.
+
+```
+Layer 6: LanceDB (cold persistence)         ← lance-graph
+Layer 5: GPU/APU meta ops (optional)        ← future
+Layer 4: Planner strategies (16-19)         ← lance-graph-planner
+Layer 3: CollapseGate write protocol        ← ndarray (enum) + contract
+Layer 2: CognitiveShader dispatch           ← p64-bridge
+Layer 1: BindSpace columns + multi-lane     ← ndarray + contract
+Layer 0: SIMD primitives                    ← ndarray (THIS CRATE)
+```
+
+## Public Surface: `ndarray::simd::*`
+
+All consumers import from `ndarray::simd::*`, NOT from `ndarray::hpc::*`.
+The hpc/ paths are private implementation detail. The simd/ module is the
+stable public API.
+
+Types that MUST be in `ndarray::simd::*`:
+- `F32x16, F64x8, U8x64, F16x32, U64x8, I16x32, I8x64`
+- `Fingerprint<N>` — const-generic, N×64 bits
+- `MultiLaneColumn<T>` — same bytes, multiple SIMD lane views
+- `array_window(data, N)` — aligned batch iterator
+- `VectorWidth, vector_config()` — the LazyLock width singleton
+- `CollapseGate` — Flow/Block/Hold enum (exists in hpc/bnn_cross_plane)
+
+If a type isn't in `ndarray::simd::*`, consumers can't use it.
+Keeps our API surface small. Internal refactors in hpc/ don't break
+downstream.
+
+## What Layer 0 Provides
+
+### SIMD Primitives (hardware abstraction)
+- Runtime dispatch: `simd_caps()` frozen singleton
+- AVX-512 (F32x16, VPOPCNTDQ, VPGATHERDD)
+- AVX2 + FMA
+- NEON (A53 / A72 / A76 dotprod tiers)
+- AMX via `asm!(".byte ...")` — TDPBF16PS, TDPBUSD
+- F16C hardware conversion
+- BF16 bit-exact RNE matching VCVTNEPS2BF16
+
+### Fingerprint<N> — the BindSpace atom
+- `[u64; N]` backing, 64-byte aligned
+- `get/set/toggle_bit`, `bind` (XOR), `and`, `not`
+- `hamming_distance` via SIMD popcount
+- `popcount`, `density`
+- `random` (xorshift128+), `from_content` (hash expansion)
+- `permute` (circular bit shift for sequence encoding)
+
+### MultiLaneColumn — same object, multiple SIMD widths
+- One `Arc<[u8]>` backing store
+- View as U8x64 / F16x32 / F32x16 / F64x8 without copy
+- Consumer picks lane width per operation
+
+### array_window — SIMD batch iterator
+- Yields N-aligned chunks from a slice
+- Zero-copy: window IS a `&[T]` view
+- One cascade level = one array_window pattern
+
+### CollapseGate enum
+- `Flow` / `Block` / `Hold` (already in `hpc/bnn_cross_plane`)
+- Consumers (L3) extend with MergeMode (Xor/Bundle/Superposition)
+
+## What ndarray DOES NOT Provide
+
+These live UP the stack, not in ndarray:
+- BindSpace address types (lance-graph-contract)
+- CognitiveShader dispatch (p64-bridge)
+- Planner strategies (lance-graph-planner)
+- CausalEdge64 (causal-edge)
+- NARS inference (causal-edge + contract)
+- GGUF parsing (bgz-tensor / consumer)
+
+Keep ndarray free of cognitive logic. It's the foundation, not the cortex.
+
+## Current Gaps (next session targets)
+
+1. **MultiLaneColumn type doesn't exist yet** — add to `src/hpc/column.rs`,
+   re-export from `src/simd.rs`
+2. **Fingerprint<N> missing `as_u8x64()`** — add SIMD view methods
+3. **simd.rs re-exports incomplete** — add Fingerprint, MultiLaneColumn,
+   array_window, VectorWidth
+4. **VectorWidth LazyLock not consumed** — any module that serializes
+   fingerprints should read it for width config
+5. **Hamming popcount hasn't been exposed via multi-lane view** —
+   combine with MultiLaneColumn for the Layer 1 cascade path
+
+## Migration Tracking (from ladybug-rs)
+
+ladybug-rs depended on `rustynum` as its HPC crate. rustynum was
+ported INTO this ndarray fork as `src/hpc/` (55 modules, 880 tests).
+Downstream consumers (lance-graph-cognitive, learning crate) still
+reference `rustynum_core::*` types. They need these substitutions:
+
+| ladybug-rs `rustynum_core::*` | ndarray equivalent |
+|---|---|
+| `Fingerprint` | `ndarray::simd::Fingerprint<256>` |
+| `hamming_distance` | `ndarray::hpc::bitwise::hamming_distance_raw` |
+| `simd_level` | `ndarray::hpc::simd_caps::simd_caps()` |
+| `cascade::*` | `ndarray::hpc::cascade::*` |
+| `bf16_*` | `ndarray::hpc::quantized::BF16` |
+| `rustynum_bnn::CollapseGate` | `ndarray::hpc::bnn_cross_plane::CollapseGate` |
+| `rustynum_holo::*` | `ndarray::hpc::holo::*` |
+
+**migration-tracker agent** owns this substitution table.
+
+## The Endgame (ndarray's view)
+
+Each token of LLM inference in the cognitive shader system runs:
+
+```
+1. Read BindSpace column slice                → &[u64; N]  (Layer 1)
+2. Hamming popcount via SIMD dispatch         → [u32; N]   (Layer 0)
+3. Base17 L1 distance on survivors            → [u16; M]   (Layer 0)
+4. Palette table lookup (256×256)             → [u8; K]    (Layer 0)
+5. Gather via VPGATHERDD                      → [u8; K]    (Layer 0)
+```
+
+All Layer 0. All ndarray. Zero FP. Zero matmul. 611M lookups/sec.
+
+The cognitive layers above coordinate WHICH columns to scan and HOW
+to combine results. ndarray just executes the primitives as fast as
+the hardware allows. Pi Zero to Sapphire Rapids, same API, same
+correctness, different throughput.
diff --git a/src/simd.rs b/src/simd.rs
index 8ccd2ff6..ae35a939 100644
--- a/src/simd.rs
+++ b/src/simd.rs
@@ -949,6 +949,42 @@ pub fn simd_ln_f32(x: F32x16) -> F32x16 {
     F32x16::from_array(out)
 }
 
+// ============================================================================
+// Cognitive shader foundation re-exports
+// ============================================================================
+
+// Fingerprint<N>: const-generic binary vector, the BindSpace atom
+pub use crate::hpc::fingerprint::{
+    Fingerprint,
+    Fingerprint2K, Fingerprint1K, Fingerprint64K,
+    VectorWidth, VectorConfig, vector_config,
+};
+
+// CollapseGate: Flow/Block/Hold write gate (Layer 3 in the 7-layer stack)
+pub use crate::hpc::bnn_cross_plane::CollapseGate;
+
+// Bitwise: SIMD-dispatched Hamming distance + popcount
+pub use crate::hpc::bitwise::{
+    hamming_distance_raw, popcount_raw,
+};
+
+// WHT: Walsh-Hadamard Transform (SIMD butterfly)
+pub use crate::hpc::fft::{wht_f32, wht_f32_new};
+
+// Quantization: i4/i2/i8 pack/unpack + BF16
+pub use crate::hpc::quantized::{
+    quantize_f32_to_i4, dequantize_i4_to_f32,
+    quantize_f32_to_i2, dequantize_i2_to_f32,
+    quantize_f32_to_i8, dequantize_i8_to_f32,
+    QuantParams,
+};
+
+// K-means + L2 distance
+pub use crate::hpc::cam_pq::{kmeans, squared_l2};
+
+// SIMD cosine
+pub use crate::hpc::heel_f64x8::cosine_f32_to_f64_simd;
+
 // ============================================================================
 // Tests
 // ============================================================================