Closes the two format gaps v1's README explicitly deferred, after
vetting real pure-Rust alternatives:
MP3 via nanomp3 (MIT/Apache-2.0): a safe-Rust port of the well-
established minimp3 C algorithm. The obvious `minimp3` crate is FFI to
that C library and its own README discloses "multiple memory
unsoundness issues" -- nanomp3 is the pure port its own README points
to as the recommended alternative. Low unsafe density (~0.6% of a
6300-line file), no_std, zero dependencies.
MP4/M4A (AAC-LC/HE-AAC) via mp4 (pure-Rust ISO-BMFF demux, MIT, mature
since 2018) + oxideav-aac (pure-Rust AAC decode, MIT). redlux -- the
most established MP4/AAC crate -- turned out to depend on fdk-aac, FFI
to a C library, disqualifying it on the same grounds as symphonia.
oxideav-aac's top-of-file doc comment is stale (claims "decode bodies
not wired up yet") but its actual decode::StreamDecoder is complete:
SBR/PS support, and a pcm_byte_exact test suite that compares decoded
output sample-for-sample against a reference decoder with an honestly-
reasoned tolerance model (<=1 LSB deterministic, energy-correctness
only for PNS noise). Demux gives per-sample raw AAC access units +
AudioSpecificConfig fields (audioObjectType/samplingFrequencyIndex/
channelConfiguration); decode_raw_data_block is oxideav-aac's
transport-independent entry point that consumes them directly (no
ADTS framing needed for MP4's raw-frame storage).
format.rs sniffs MP4 via the ftyp box signature at offset 4, MP3 via
either a leading ID3v2 tag or a raw MPEG frame sync (0xFF + top 3 bits
set).
Verified: 8 new tests (5 mp3, 5 mp4 unit tests across
clawaudio-decode, minus overlap with sniff tests; 2 new end-to-end
facade tests) against real ffmpeg-generated MP3/M4A fixtures --
decode -> resample -> spectrogram all non-degenerate, dominant
frequency recovered within lossy-codec tolerance. Full workspace
fmt/clippy -D warnings clean.
Co-Authored-By: Claude Sonnet 5 <[email protected]>
