//! # `FermionResearch`'s model formats
//!
//! The Phonon speech models ship in formats of their own. Every one is read
//! here to plain f32 values, exactly as the format's reference reader
//! expands it, and kits build on the results:
//!
//! - `PhononArchive` (feature `fermion`): the release archive of every Phonon
//! model, a zstd tar with a digest-checked manifest whose safetensors members
//! may be stored byte-plane split.
//! - [`read_five_value_container`]: the `fermion-five-value-parakeet-v1`
//! container inside Phonon-2's archive, the weights of
//! [`kits::speech::parakeet`](crate::kits::speech::parakeet).
//! - [`decode_quint5`]: the `ten-base5-per-24bit-v1` codes Phonon-1 stores its
//! text-decoder linears in, inside safetensors members whose other records
//! are bf16 or MLX affine-quantized
//! ([`dequantize_mlx_affine`](super::mlx::dequantize_mlx_affine)).
//!
//! The fixtures in `testdata/fermion/` were written by
//! `tools/gen_fermion_fixtures.py` with the formats' own readers and writer.
pub use *;
pub use *;
pub use *;