Module codec 

Codec layer: deterministic CodecConfig, canonical rotation matrix, rotation cache, codebook quantization, residual helpers, the stateless Codec service, and the PreparedCodec session object.

Phase 13 introduced the rotation primitives; Phase 14 adds the uniform-quantile Codebook plus scalar quantize / dequantize kernels. Phase 15 adds the Codec service, FP16 residual helpers, CompressedVector value object, and Parallelism enum. Phase 26 adds PreparedCodec — a pre-built session that caches the rotation matrix for reuse across batch calls, enabling the GPU backend attachment point (Phase 27+).

Re-exports

pub use codebook::Codebook;

pub use codec_config::CodecConfig;

pub use codec_config::MAX_DIMENSION;

pub use codec_config::SUPPORTED_BIT_WIDTHS;

pub use compressed_vector::CompressedVector;

pub use parallelism::Parallelism;

pub use prepared::PreparedCodec;

pub use rotation_cache::RotationCache;

pub use rotation_cache::DEFAULT_CAPACITY as ROTATION_CACHE_DEFAULT_CAPACITY;

pub use rotation_matrix::RotationMatrix;

pub use service::compress;

pub use service::decompress;

pub use service::Codec;

Modules

codebook

Immutable quantization lookup table with Python-parity training.

codec_config

CodecConfig: immutable value object describing codec parameters.

compressed_vector

In-memory CompressedVector value object (Phase 15).

gaussian

Deterministic ChaCha20-backed standard-normal f64 stream.

parallelism

Parallelism dispatch for batch codec methods (Phase 15).

prepared

PreparedCodec: pre-built session object for the hot compress/decompress path.

residual

FP16 residual helpers (Phase 15).

rotation_cache

Process-wide LRU cache for rotation matrices, keyed by (seed, dim).

rotation_matrix

Canonical orthogonal rotation matrix (ChaCha → faer QR → sign correction).

service

Stateless Codec service (Phase 15).