fix(deps): surgical hpc-extras gate — only p64/fractal behind feature

Before: hpc-extras gated the entire `pub mod hpc;`, hiding BF16/F16,
quantization, fingerprints, WHT, kmeans, reductions, AMX, VNNI, NEON,
VSA, plane, seal — everything burn-ndarray actually needs.

After: `pub mod hpc;` is always available with `std`. Only 5 truly
specialized research modules gated behind hpc-extras:
- p64_bridge (Palette64 convergence highway)
- crystal_encoder (blake3-hashed codebook)
- deepnsm (distributional semantic engine)
- spo_bundle (SPO triple bundle codec)
- compression_curves (blake3 XOF binary planes)

blake3 made unconditional (was optional) — it's core infrastructure
for plane/seal/merkle_tree integrity hashing, not research-only.

Verified:
- cargo build --no-default-features --features std: PASS
- cargo build (defaults): PASS
- cargo test (no-hpc-extras): 1713 pass
- cargo test (defaults): 1767 pass

https://claude.ai/code/session_01NYGrxVopyszZYgLBxe4hgj

Author: claude

Cargo.toml

@@ -49,10 +49,11 @@ libc = { version = "0.2.82", optional = true }
 
 matrixmultiply = { version = "0.3.2", default-features = false, features=["cgemm"] }
 
-# HPC extras — gated behind the `hpc-extras` feature so downstream consumers
-# (e.g. burn-ndarray) can opt out of the heavy compile tree when only the core
-# array layer is needed. Enabled by default to preserve existing behavior.
-blake3 = { version = "1", optional = true }
+# blake3 — always available (integrity hashing in plane/seal/merkle_tree).
+blake3 = { version = "1" }
+
+# p64 + fractal — specialized convergence / manifold math. Gated behind
+# `hpc-extras` since they pull in a dep tree burn-ndarray doesn't need.
 p64 = { path = "crates/p64", optional = true }
 fractal = { path = "crates/fractal", default-features = false, optional = true }
 
@@ -92,7 +93,7 @@ rayon = ["dep:rayon", "std"]
 # These pull in a non-trivial dependency tree; downstream crates such as
 # burn-ndarray that only need the core array layer can disable this with
 # `default-features = false` (and re-enable `std` explicitly if needed).
-hpc-extras = ["std", "dep:blake3", "dep:p64", "dep:fractal", "fractal/std"]
+hpc-extras = ["std", "dep:p64", "dep:fractal", "fractal/std"]
 
 matrixmultiply-threading = ["matrixmultiply/threading"]
 

src/hpc/mod.rs

@@ -114,22 +114,27 @@ pub mod tekamolo;
 #[allow(missing_docs)]
 pub mod vsa;
 #[allow(missing_docs)]
+#[cfg(feature = "hpc-extras")]
 pub mod spo_bundle;
 #[allow(missing_docs)]
+#[cfg(feature = "hpc-extras")]
 pub mod deepnsm;
 #[allow(missing_docs)]
 pub mod surround_metadata;
 #[allow(missing_docs, dead_code)]
 pub mod cyclic_bundle;
 #[allow(missing_docs, dead_code)]
+#[cfg(feature = "hpc-extras")]
 pub mod compression_curves;
 #[allow(missing_docs)]
+#[cfg(feature = "hpc-extras")]
 pub mod crystal_encoder;
 #[allow(missing_docs)]
 pub mod udf_kernels;
 
 // p64 bridge: Palette64/3D attention, NARS, CausalEdge64 compat
 #[allow(missing_docs)]
+#[cfg(feature = "hpc-extras")]
 pub mod p64_bridge;
 
 // Session C: bgz17 dual-path integration
@@ -228,7 +233,7 @@ pub mod framebuffer;
 /// Transcoded from Opus CELT for the HHTL cascade → waveform pipeline.
 pub mod audio;
 
-#[cfg(test)]
+#[cfg(all(test, feature = "hpc-extras"))]
 mod e2e_tests {
     //! End-to-end pipeline test: Fingerprint → Node → Seal → Cascade → CLAM → Causality → BNN
 

src/lib.rs

@@ -258,6 +258,7 @@ pub mod simd_wasm;
 pub mod simd_int_ops;
 
 /// Half-precision SIMD vectors (`BF16x16`, `F16x16`) + slice-level ops.
+/// Depends on `hpc::quantized::{BF16, F16}` — needs `std` (hpc core).
 #[cfg(feature = "std")]
 #[allow(clippy::all, missing_docs, dead_code, unused_variables, unused_imports)]
 pub mod simd_half;
@@ -268,10 +269,11 @@ pub mod backend;
 
 /// HPC extensions ported from rustynum: BLAS, statistics, HDC, CogRecord, FFT, LAPACK.
 ///
-/// Gated behind the `hpc-extras` feature (enabled by default) because the
-/// module pulls in `blake3`, `p64`, and `fractal`. Disable default features to
-/// drop those dependencies (e.g. burn-ndarray's polyfill-only build).
-#[cfg(feature = "hpc-extras")]
+/// Always available with `std`. Core modules (quantized, fingerprint, fft,
+/// cam_pq, reductions, blas_level*, amx_matmul, vnni_gemm) compile without
+/// extra deps. Cognitive/research modules (p64_bridge, crystal_encoder,
+/// deepnsm, etc.) are gated behind `hpc-extras` inside `hpc/mod.rs`.
+#[cfg(feature = "std")]
 #[allow(
     clippy::all,
     unused_imports,

src/simd.rs

@@ -1184,25 +1184,20 @@ pub fn simd_ln_f32(x: F32x16) -> F32x16 {
 // Cognitive shader foundation re-exports
 // ============================================================================
 
-// Fingerprint<N>: const-generic binary vector, the BindSpace atom
+// HPC re-exports — only available when the hpc module is compiled.
+// Without `hpc-extras`, consumers still get the SIMD polyfill types above
+// (F32x16, I8x32, etc.) but NOT the domain-specific functions below.
+
 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,
@@ -1234,9 +1229,11 @@ pub use crate::simd_half::{
 pub use crate::simd_half::BF16x16 as BF16x16;
 
 // 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;
 
 // ============================================================================