feat(hpc): Fingerprint BindSpace API + VectorWidth config + WHT + BF16 tile GEMM + i2 quantization

[AdaWorldAPI]

Summary

Extends ndarray with the hardware + type primitives that lance-graph's
cognitive-shader-driver consumes. Everything in the contract crate depends
on these directly (Fingerprint, VectorWidth, SIMD lane views, BLAS-adjacent
kernels, quantization helpers).

src/hpc/fingerprint.rs (+236 lines)

Full BindSpace-compatible API on Fingerprint<N>:

• Bit ops: get_bit, set_bit, toggle_bit
• Algebra: bind (XOR), and, or, not, permute
• Constructors: random(seed), orthogonal(seed), from_content(&str)
• Stats: density, hamming (alias)
• Bundling: bundle(items: &[&Self]) — majority vote
• SIMD views: chunks_u64x8, chunks_u8x64 — zero-copy lane iteration
• Width config: VectorWidth enum + LazyLock singleton + vector_config()
  reading NDARRAY_VECTOR_WIDTH env var (production 16K default)

Six new types are now part of the public surface via simd re-exports.

src/hpc/quantized.rs (+48 lines)

• quantize_f32_to_i2 / dequantize_i2_to_f32 — 2-bit precision for the
  cascade path
• dequantize_i8_to_f32 — paired reverse for the existing i8 codec
• QuantParams public

src/hpc/fft.rs (+135 lines)

• wht_f32(&mut [f32]) — Walsh–Hadamard Transform with F32x16 SIMD butterfly
• wht_f32_new(&[f32]) — functional variant

Used by the cognitive shader's HAD-cascade codec.

src/hpc/bf16_tile_gemm.rs (+198 lines) + src/hpc/amx_matmul.rs (+44 lines)

• bf16_tile_gemm — AMX TDPBF16PS primitive with AVX-512 polyfill
• 16×16 tile matrix multiply for BF16 × BF16 → f32 accumulation
• Runtime dispatch through simd_caps()

src/simd.rs (+36 lines)

Public re-exports for lance-graph consumers:

pub use crate::hpc::fingerprint::{
    Fingerprint, Fingerprint2K, Fingerprint1K, Fingerprint64K,
    VectorWidth, VectorConfig, vector_config,
};
pub use crate::hpc::bnn_cross_plane::CollapseGate;
pub use crate::hpc::bitwise::{hamming_distance_raw, popcount_raw};
pub use crate::hpc::fft::{wht_f32, wht_f32_new};
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,
};
pub use crate::hpc::cam_pq::{kmeans, squared_l2};
pub use crate::hpc::heel_f64x8::cosine_f32_to_f64_simd;

Consumers write use ndarray::simd::{Fingerprint, VectorWidth, ...};
and never touch internal hpc::* paths.

.claude/knowledge/cognitive-shader-foundation.md (+137 lines)

Agent knowledge doc parallel to lance-graph's. Explains the SIMD floor
(F32x16), the 4-tier dispatch (F32x16 → VNNI2 → AVX512-VNNI → AMX),
the Fingerprint const-generic model, the VectorWidth LazyLock
config path, and which public types lance-graph consumes.

.claude/agents/*.md model bumps

Four agents (l3-strategist, migration-tracker, product-engineer,
vector-synthesis) updated to the Opus 4.7 model tag.

Test plan

• cargo test --lib fingerprint — 21 passing
• cargo check — clean
• Existing ndarray tests unaffected (1639 filtered-out tests in the
  pattern-matched run are from other modules, all passing in full cargo test)
• Downstream consumer verified: cargo test -p lance-graph-contract
  and cargo test -p cognitive-shader-driver compile and pass with
  these additions (tested during lance-graph PR #206)

Downstream impact

lance-graph's cognitive-shader-driver and lance-graph-contract both
import from ndarray::simd::* — this PR is what lets their PR #206
compile. Merging unblocks the Tier 0 quick wins (Q2 Cargo.toml pin,
AriGraph wiring, cockpit endpoints).

https://claude.ai/code/session_01SbYsmmbPf9YQuYbHZN52Zh

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 3c77050c7f

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

[chatgpt-codex-connector]

Return only full SIMD lanes in chunk iterators

chunks_u64x8 uses .chunks(8), so Fingerprint<N> values where N % 8 != 0 produce a final slice shorter than 8 words; callers following the documented U64x8::from_slice(chunk) usage will panic on that tail chunk. Because Fingerprint is a public const-generic type (and this file already uses small non-multiple test sizes), this API can crash on valid inputs unless it uses chunks_exact(8) (or otherwise handles the remainder explicitly).

Useful? React with 👍 / 👎.

[chatgpt-codex-connector]

Prevent orthogonal(0) from collapsing to zero fingerprint

orthogonal forwards the transformed seed into random; when seed == 0, this still passes 0, and the xorshift state remains all-zero, yielding an all-zero fingerprint every time. That breaks the method’s stated quasi-orthogonal behavior and creates a degenerate vector if callers generate seeds starting at 0 (a common indexing pattern), so zero should be remapped/mixed to a non-zero RNG state.

Useful? React with 👍 / 👎.