Decoder SIMD dispatch

The constructor validates and stores a backend. Private decompressor::core dispatches before command loops and bulk copy stages, passing a concrete fearless_simd::Simd token into generic helpers. Headers, raw blocks and resumable parsing stages share scalar control flow.

The SIMD choice is the innermost of three compile-time selections. Stream::run first fixes the sink type and whether the member decodes into linear slice history (run_stages::<O, LINEAR>). Each of those state machines then dispatches the backend at its command loop and copy stages, so every backend is compiled once per instantiation.

flowchart LR
    Run[Stream::run] --> Machine[run_stages per sink type and LINEAR]
    Machine --> Backend[validated Backend]
    Backend[validated Backend] --> Dispatch[dispatch outside command or copy loop]
    Dispatch --> Control[serial Huffman and command dependencies]
    Dispatch --> Copy[16/32-byte array snapshots]
    Control --> State[shared output, progress, limits and errors]
    Copy --> State

The dispatch gives the command loop its backend’s target features: the bit operations compile to BMI2 shifts and masks where available, and the short copies are fixed-size array moves (copy_chunk) that those features turn into one or two vector loads and stores. Copies load the complete source before storing and accept unaligned initialized slices. One test of the later window against the last start that fits covers both windows; a short boundary uses an exact-length copy. Speculative writes remain within at most fifteen unreachable ring bytes. Long overlapping expansions use fill, byte copying or prefix doubling. Feature checks and virtual calls do not occur per byte or command.

Native decompressor specifies wrap, overlap, dictionary and error invariants. Owned output specifies history growth and transfer; those ownership rules are shared across backends.

Verification and known gaps

Unit tests use a bytewise ring oracle and snapshot-copy cases. Differential tests explicitly exercise fallback and every backend available on the host, checking payload, consumed input, errors, limits and streaming progress.

Run the root decompress Criterion benchmark for retained slices, streaming and backend comparisons; the separate library comparison measures cold owned APIs. Profile with hotpath before changing these kernels.

Huffman construction and context-map transforms remain scalar. Backend tests cannot validate instruction sets absent from their execution host; compilation alone does not establish runtime correctness or speed.