v0.0.1

Other

• SFZ + DLS voice generators (task #410)
• DLS Level 1 + 2 RIFF reader (parse + dump bank)
• SFZ text patch reader (load + dump regions)

Round 9 — SFZ + DLS voice generators (task #410)

• Shared sample-playback voice (instruments::sample_voice). Mono
  in, mono out; the mixer handles stereo panning.
  Covers the DAHDSR amplitude envelope (delay / attack / hold / decay /
  sustain / release), four loop modes (NoLoop, OneShot,
  LoopContinuous, LoopSustain), pitch bend via the existing
  Voice::set_pitch_bend_cents hook, channel/poly aftertouch via
  Voice::set_pressure, exclusive-class drum cuts, and a sine vibrato
  LFO with rate / depth / start-delay.
• Minimal RIFF/WAVE PCM decoder (instruments::wav_pcm). Decodes
  8-bit unsigned, 16-bit signed LE, 24-bit signed LE, 32-bit signed LE
  PCM, and 32-bit IEEE_FLOAT into mono f32. Stereo / multi-channel WAVs
  are mixed down to mono by averaging channels (round-2 voice
  generation will keep stereo intact).
• SFZ voice generator. SfzInstrument::make_voice walks the
  flattened region table for the highest-priority match on (key, velocity), decodes the WAV bytes loaded by SfzInstrument::open,
  shifts pitch off pitch_keycenter + tune + transpose, and
  instantiates a SamplePlayer honoring the region's loop_* opcodes
  • an amplitude envelope from ampeg_delay/attack/hold/decay/sustain/ release + a vibrato LFO from lfo01_freq / lfo01_pitch /
    lfo01_delay (with vibrato_* aliases).
• DLS Level 1 + 2 voice generator. DlsInstrument::make_voice
  picks the matching instrument by MIDI program (bank-MSB / LSB
  matching is round 2), picks a region by (key, velocity), resolves
  wlnk.table_index → ptbl cue → wave-pool entry, decodes the PCM
  via wav_pcm::decode_pcm_bytes, and plays the sample through the
  shared SamplePlayer. Region-level wsmp overrides the wave-level
  default per the spec; WLOOP_TYPE_FORWARD (0) maps to
  LoopContinuous, WLOOP_TYPE_RELEASE (1) maps to LoopSustain.
  art1/art2 connection-block evaluation is round 2 — the parsed
  blocks remain on the bank.
• InstrumentSource builder + MidiDecoder::with_instrument_source.
  Caller passes InstrumentSource::sf2(path) / sfz(path) / dls(path)
  / Tone and the decoder picks the right loader. Format detection is
  not by extension — the caller picks the variant.
• Tests added: 13 net new lib-side (5 sample_voice, 4 wav_pcm, 3
  SFZ voice-generation, 2 DLS voice-generation, minus the 2 existing
  make_voice_returns_unsupported tests that were replaced/upgraded
  to actually exercise the round-1 voice path) + 3 integration
  (tests/voice_round_trip.rs) exercising end-to-end SFZ/SF2/DLS
  rendering through MidiDecoder with an RMS non-silence assertion.
  Total: 156 lib + 6 integration = 162 passing (was 143 + 3 = 146).

Round 8 — DLS Level 1 + 2 sample-loader (task #409)

• DLS RIFF parser: walks the RIFF/DLS form and pulls the
  colh collection header, optional vers version stamp, ptbl
  pool table, lins instrument list, wvpl wave pool, and
  top-level INFO metadata into a fully-resolved
  DlsBank. Instruments → regions →
  wave-pool samples are cross-referenced; nothing references back
  into the source bytes.
• Wave pool: every wave-list entry is parsed for its standard
  WAV fmt + data chunks plus the optional wsmp per-wave loop
  / pitch / gain header. Sample bytes are kept in their on-disk
  form (8-bit unsigned or 16-bit LE signed); decode is round-2.
• Instrument table: each ins LIST surfaces its bank/program
  (decoded into bank_msb / bank_lsb / program_number and the
  is_drum() bit-31 helper), instrument name from a per-instrument
  INFO/INAM, and an instrument-level articulation list parsed
  from lart (DLS1) or lar2 (DLS2) sub-LISTs.
• Regions: rgnh (key + velocity range, fusOptions, key group,
  optional DLS2 usLayer), wsmp (per-region overrides), wlnk
  (cue-table reference), and per-region articulation. DLS2 rgn2
  LISTs parse alongside DLS1 rgn and are flagged via
  DlsRegion::is_level2.
• Articulation: art1-ck and art2-ck connection blocks
  (12-byte records: source / control / destination / transform /
  scale) parse into Vec<DlsArticulationBlock> tagged with
  DlsArtKind::{Art1, Art2} so the round-2 voice generator picks
  the right enum table (DLS1 spec page 43 / DLS2 spec tables 8-10).
  Connection enums are stored as raw u16s — no interpretation in
  round 1.
• Magic-byte stub becomes real probe + parser: is_dls() and
  the new DlsInstrument::probe() honour the RIFF/DLS magic;
  DlsInstrument::open() and parse_bytes() plumb through to the
  full bank parser. make_voice() still returns
  Error::Unsupported (round-2 work, same shape as the SFZ
  followup).
• Bounds + caps: every chunk length is checked against bytes
  remaining; pool-table, articulation, and wave-pool counts are
  capped at MAX_RECORDS (1 Mi); cumulative wave-data bytes capped
  at MAX_WAVE_BYTES (256 MiB).
• Tests added: 13 lib-side (magic detection, minimal-DLS
  parse + wave pool + instrument + region + articulation, DLS2
  rgnh-with-usLayer, art2 block, wsmp loop record, error paths
  for non-DLS / truncated outer / non-DLS path, drum-bit decode,
  open round-trip through disk) + 1 integration smoke
  (tests/sfz_sf2_dls_smoke.rs) building a 2-region DLS in
  memory and dumping the instrument + region table. Total: 143
  lib + 3 integration = 146 passing (was 130 + 2 = 132).
• Smoke test renamed from tests/sfz_sf2_smoke.rs to
  tests/sfz_sf2_dls_smoke.rs to reflect the wider coverage.

Round 7 — SFZ text patch reader (task #127)

• SFZ parser: tokenises SFZ syntax (line // ... + block
  /* ... */ comments, <header> sections, name=value opcode
  pairs with space-bearing values like sample paths) and walks the
  full <control> / <global> / <master> / <group> / <region>
  hierarchy. Inheritance is flattened into one fully-resolved opcode
  map per region (global → master → group → region, later overrides
  earlier).
• Strongly-typed region fields for the round-2 voice generator:
  sample_path, lokey / hikey / lovel / hivel,
  pitch_keycenter, key (sets lokey + hikey + pitch_keycenter),
  loop_start / loop_end, loop_mode (no_loop / one_shot /
  loop_continuous / loop_sustain), transpose, tune (alias
  pitch), volume, pan, trigger. Note names (C4, c#4,
  Db5, c-1) parse alongside decimal MIDI keys.
• Sample loader: SfzInstrument::open resolves every sample=
  path against the SFZ file's directory + the active <control> default_path= opcode and reads the bytes off disk into
  region.sample_bytes. Missing or unreadable samples become a hard
  parse error so the caller learns at load time. parse_str skips
  the filesystem hooks for in-memory tests.
• Preprocessor: #include is rejected with Error::Unsupported
  (round-1 reader doesn't follow includes); #define is stored
  verbatim in the surrounding scope's opcode map without macro
  expansion.
• DLS reader status: docs-blocked. The new
  docs/audio/midi/instrument-formats/ directory contains the SFZ
  format docs (10 HTML files) plus the SoundFont 2.04 spec PDF, but
  no Microsoft DLS Level 1/2 specification. The DLS magic-byte stub
  remains in place; voice generation continues to return
  Error::Unsupported.
• New tests: 23 added — 22 lib-side covering tokenisation, comment
  stripping, key-name parsing, header inheritance, group reset,
  control / default_path resolution, loop opcodes, opcode-map
  preservation, #include rejection, sample loading + missing-file
  handling, and a tutorial-shaped template smoke; 2 integration tests
  (tests/sfz_sf2_smoke.rs) that dump SFZ regions + an SF2 preset
  list via the public API. Total: 130 lib + 2 integration = 132
  passing (was 111).

Round 6 — SF2 polish (sm24 + stereo + mod-env + filter, task #139)

• 24-bit sample storage (SF2 2.04+ sm24 chunk). PCM is now
  stored as Arc<[i32]> carrying signed 24-bit values in the lower
  24 bits. When sm24 is present its u8 lower bytes are combined
  with the 16-bit smpl upper bytes; otherwise the 16-bit value is
  widened by left-shift-8. Mismatched sm24 length is silently
  ignored per spec ("parsers must tolerate"). Voice fetch divides
  by 2^23 instead of 2^15.
• Stereo SF2 zones. Sample headers tagged LEFT / RIGHT with
  a valid bidirectional sample_link are detected at resolve time
  and produce a stereo-aware Sf2Voice that holds two phase
  counters and writes distinct L/R via the new Voice::render_stereo
  hook. The mixer routes such voices through a balance law (cos/sin
  scaled to unity at centre) rather than its mono-pan path.
• Modulation envelope (gens 25-30 — delay/attack/hold/decay/
  sustain/release). Same DAHDSR shape as the volume envelope but
  with 0..=1 sustain levels; release tracks the volume envelope's
  release_pos so a note-off cleanly tails both off together.
• Mod-env routing (gens 7 + 11). modEnvToPitch adds the
  envelope-scaled cents offset to the live pitch-bend cents on every
  sample. modEnvToFilterFc modulates the biquad cutoff in cents.
• Initial low-pass filter (gens 8/9). Direct-form-1 RBJ-cookbook
  biquad on the voice output. Cutoff in absolute cents (re. 8.176
  Hz), Q in centibels of resonance. Filter state is allocated only
  when the cutoff is below ~12 kHz or the mod-env routes meaningfully
  to it; bypass otherwise.
• Exclusive class (gen 57). Note-on with the same non-zero class
  on the same channel hard-stops every prior voice in that class —
  used for hi-hat open/closed pairs in drum kits. Implemented in
  the mixer via a new Voice::exclusive_class hook.
• Pitch-wheel range RPN 0 verified end-to-end. ChannelState:: pitch_bend_range_cents defaults to 200 (±2 semitones); CC 100/101
  selects RPN 0; CC 6/38 sets the semitone+cent range. Live bend is
  re-applied on range change so still-held voices pick up the new
  scale. Existing tests cover the path.
• New tests: 16 added (sm24 combine + missing-sm24 fallback +
  wrong-length tolerance; stereo resolve + render + mixer routing +
  self-link rejection; filter HF attenuation + default bypass;
  mod-env routing + brightening; exclusive class propagate + cut;
  overridingRootKey verification; 24-bit grid sanity). Total: 111
  passing (was 95).