v0.0.4

Other

• midi r307: SmfFile::active_notes_at() sounding-note seek lens
• add SmfFile::poly_aftertouches() per-key aftertouch iterator
• add SmfFile::notes() Note On/Off pairing into sounding-note spans
• run-segmented volume-envelope evaluation — bit-identical, ~20% faster SMF→PCM synthesis
• add SmfFile::channel_pressures() — Dn-pp mono-aftertouch iteration helper
• SMF SmfFile::pitch_bends() — En-lsb-msb channel-voice pitch-bend iteration helper
• SmfFile::control_changes() — Bn-cc-vv channel-voice continuous-controller / channel-mode iteration helper
• SmfFile::program_changes() — Cn-pp channel-voice patch-select iteration helper
• SmfFile::universal_sysex_events() — Table-4-classified file-wide iteration helper
• SysExEvent::universal_classification() — Table 4 Universal SysEx classifier
• SmfFile::sysex_events() iteration helper (F0 / F7)
• SmfFile::channel_prefixes() iteration helper (FF 20 01 cc)
• SmfFile::to_bytes() / Track::to_bytes_chunk() mux-side writer
• SmfFile::midi_ports() iteration helper (FF 21 01 pp)
• SmfFile::sequence_numbers() iteration helper (FF 00 02 ssss)
• SmfFile::sequencer_specifics() iteration helper (FF 7F)
• SmfFile::channel_snapshot_at / channel_snapshots_at (channel-state seek primitive)
• SmfFile::smpte_offsets() iteration helper + FrameRate decoder (FF 54)
• SmfFile::texts() + copyrights() iteration helpers (FF 01 + FF 02)
• SmfFile::instrument_names() iteration helper (FF 04)

Round 307 — SmfFile::active_notes_at() sounding-note seek lens

• New SmfFile::active_notes_at(tick) -> Vec<Note>: every Note span
  sounding at the absolute tick — the piano-roll / seek companion to
  notes(), and the note-level analogue of the channel-state
  channel_snapshot_at primitive. Where the snapshot answers "what
  controller / program / bend state does a channel carry at tick T?",
  this answers "which keys are held down at tick T?" — exactly the set a
  DAW must re-trigger (or a renderer must prime into the voice pool) when
  seeking into the middle of a file rather than playing from the top.
• A note is sounding when start_tick <= tick && end_tick > tick — the
  half-open interval [start_tick, end_tick). The onset tick is
  inclusive (the snapshot reflects state immediately after that tick's
  events fire, the same convention as channel_snapshot_at); the release
  tick is exclusive (the key has come up). A zero-duration note
  (start_tick == end_tick) is sounding at no tick. Hanging onsets and
  unmatched releases — already dropped by notes() — cannot be reported.
  The result preserves the notes() (start_tick, track) order so chord
  notes stay grouped and track 0 precedes track 1.
• Seven new unit tests: empty-when-silent, half-open boundary
  (onset-inclusive / release-exclusive / mid-span / after-release),
  before-onset silence, zero-duration never sounds, chord returns all
  held keys in onset order, staggered overlap window (only-n1 / both /
  only-n2), and hanging-note never sounds. Full lib suite 550 → 557
  tests, zero ignored.

Round 301 — SmfFile::poly_aftertouches() Polyphonic Key Pressure stream

• New SmfFile::poly_aftertouches() -> Vec<PolyAftertouchEvent>: every
  An kk pp Polyphonic Key Pressure (per-key aftertouch) channel-voice
  event as a PolyAftertouchEvent { tick, track, channel, key, pressure } pinned to the absolute tick on its parent track, stably merged
  across tracks (track 0 before track 1 at the same tick) under the same
  convention every existing iteration helper and the scheduler use.
• An was the only channel-voice status nibble without a dedicated
  typed extraction helper. Distinct from Channel Pressure (Dn,
  surfaced by channel_pressures()) — An carries a per-key kk byte,
  so per-voice aftertouch automation can be rebuilt. Accessors
  channel() / key() / pressure() return the decoded fields.
• Nine new unit tests: tick-zero decode, low-nibble channel index,
  running-status (key, pressure) pair chaining, late-position absolute
  tick, two-track stable sort, cross-track tick merge, filter exclusion
  of every other channel-voice kind, to_bytes()/parse() round trip,
  and empty-when-none. Full lib suite 541 → 550 tests, zero ignored.

Round 292 — SmfFile::notes() Note On / Note Off pairing into sounding-note spans

• New SmfFile::notes() -> Vec<Note>: pairs every Note On
  (9n key vel, vel > 0) with the Note Off that releases it and
  returns one Note { start_tick, end_tick, track, channel, key, velocity, off_velocity } span per sounding note, ordered by onset.
  Where the channel-voice helpers (program_changes /
  control_changes / pitch_bends / channel_pressures) surface one
  value per wire event, this helper joins the two wire events that
  bracket a note into a single span carrying its duration — the
  primitive a piano-roll / DAW note-lane view consumes directly.
• Honours the MIDI 1.0 Summary of MIDI Messages Table 1 velocity-0
  convention: a 9n key 0 is treated as a Note Off and closes the
  earliest open note of that pitch (FIFO), with off_velocity == 0.
  An explicit 8n key off_vel carries its release velocity through to
  Note::off_velocity.
• Matches over the globally merged event stream sorted by
  (absolute tick, track, in-track position) — the same stable-merge
  convention every other iteration helper and the scheduler use — so a
  Note Off on a different track from its Note On still pairs correctly.
  Overlapping notes of the same (channel, key) are matched FIFO; an
  unmatched release or a hanging onset is dropped from the span list.
• Note accessors: channel() / key() / velocity() /
  off_velocity() / duration_ticks() (end_tick - start_tick).
• 13 new unit tests: single-note pairing, velocity-0 close, channel
  decode, FIFO overlap, chord (distinct pitches + cross-track
  ordering), cross-track off pairing, hanging-on / unmatched-off drop,
  zero-duration span, sibling-helper isolation, and a to_bytes() /
  parse structural round trip.

Round 285 — synthesis profiling + run-segmented volume envelope (bit-identical, ~20 % faster)

• New benches/synth_render.rs (harness = false) — repeatable
  SMF→PCM wall-clock harness: dense 8-channel / 32-voice score with
  pitch-bend sweeps + volume/pan CCs rendered through an in-memory
  looping SF2 bank. --corpus hashes (FNV-1a-64) the PCM for every
  in-tree fixture SMF through both the SF2 bank and the tone
  fallback; --spin SECS loops the render as a sampling-profiler
  target; default mode prints per-iteration wall time + output hash.
• Profiling ranked Sf2Voice::render at ~89 % of the synthesis wall
  clock; within it the per-sample DAHDSR volume-envelope stage walk
  (an Option test + up to four stage comparisons + an f32 divide,
  serialised behind the phase walk) at ~31 % of total, ahead of
  sample fetch + linear interpolation (~15 %).
• Sf2Voice::render now evaluates the volume envelope in
  stage-segmented runs (envelope_run) into a 256-entry stack
  buffer: delay / hold / sustain become slice fills, attack / decay /
  release become element-wise loops with no loop-carried dependency
  that the compiler vectorises. Every per-sample expression is kept
  verbatim from envelope_at, so the rendered PCM is bit-identical —
  corpus hashes are unchanged and the new
  envelope_run_matches_envelope_at_per_sample test pins
  to_bits() equality across stage boundaries, the release tail,
  and the elapsed-wrap fallback. Dense-score render: 80.2 ms →
  64.2 ms (-20 %) on an Apple-silicon dev box.

Round 275 — SmfFile::channel_pressures() — Dn-pp channel-voice mono-aftertouch iteration helper

• New SmfFile::channel_pressures(&self) -> Vec<ChannelPressureEvent>
  surfaces every Dn pp Channel Pressure (mono aftertouch)
  channel-voice event on every track, pinned to the absolute tick at
  which it fires, in time order. Each entry is a
  ChannelPressureEvent { tick, track, channel, pressure } with the
  status nibble's low four bits decoded into the spec's 0..=15
  channel index and the single data byte pp (0..=127) carrying
  "the single greatest pressure value (of all the current depressed
  keys)" per the MIDI 1.0 Summary of MIDI Messages Table 1.
• The new ChannelPressureEvent struct exposes channel() /
  pressure() accessors. The helper stays routing-agnostic — the
  pressure value's musical effect (volume / vibrato depth / filter
  cutoff) is left to the receiving instrument.
• Only Dn is selected; polyphonic key pressure (An, per-key) keeps
  its own surface, and the neighbouring CC (Bn) / program (Cn) /
  pitch-bend (En) / note (8n / 9n) channel-voice events stay
  isolated. Per-track sequences are stably merged by absolute tick
  (track 0 before track 1 at the same tick), the same convention as
  every meta-event, SysEx, and channel-voice helper.
• 8 new unit tests cover tick-zero decode, channel-index nibble,
  running-status chains (single-data-byte status), late-position
  absolute tick, stable same-tick cross-track sort, cross-track tick
  merge, cross-kind filtering, and a to_bytes() / parse round trip.

Round 267 — SmfFile::pitch_bends() — En-lsb-msb channel-voice pitch-bend iteration helper

• New SmfFile::pitch_bends(&self) -> Vec<PitchBendEvent> surfaces
  every En lsb msb Pitch Bend channel-voice event on every track,
  pinned to the absolute tick at which it fires, in time order. Each
  entry is a PitchBendEvent { tick, track, channel, value } with the
  status nibble's low four bits decoded into the spec's 0..=15
  channel index and the two data bytes combined into the 14-bit code
  (msb << 7) | lsb, 0..=0x3FFF, no-bend centre 0x2000 (the parser
  assembles the value at decode time).
• The new PitchBendEvent struct carries the same tick / track
  pair the existing channel-voice iteration helpers use plus the
  decoded channel and the assembled 14-...

v0.0.3

Other

• SmfFile::track_names() iteration helper (FF 03)
• SmfFile::cue_points() iteration helper (FF 07)
• SmfFile::lyrics() iteration helper (FF 05)
• SmfFile::markers() iteration helper (FF 06)
• add cargo-fuzz harness over smf / sf2 / dls / sfz parsers (round 172)
• fix parse_key integer overflow on huge octave magnitudes
• add SmfFile::key_signatures() iteration helper
• SmfFile::tempo_map() iteration helper (FF 51)
• SmfFile::time_signatures() iteration helper (FF 58)

Round 192 — SmfFile::track_names() iteration helper (FF 03)

• New smf::TrackNameEvent { tick, track, text } plus
  SmfFile::track_names() -> Vec<TrackNameEvent>. Collects every
  track-name meta event (FF 03 len text, the DAW-track-list
  convention from the Standard MIDI File 1.0 specification) from
  every track, pins each one to the absolute tick of its parent
  track via cumulative TrackEvent::delta sums, then merges the
  per-track sequences with a stable sort by tick — track 0 wins
  over track 1 at the same tick, matching the same merge rule used
  by SmfFile::cue_points() / SmfFile::markers() /
  SmfFile::lyrics() / SmfFile::tempo_map() /
  SmfFile::time_signatures() / SmfFile::key_signatures() and by
  scheduler.rs §"merged event list, sorted by absolute tick".
• Only FF 03 is selected. Other text-kind meta events
  (FF 01 general text, FF 02 copyright, FF 04 instrument
  name, FF 05 lyric, FF 06 marker, FF 07 cue point) are
  filtered out so callers populating a DAW track-list label don't
  have to discriminate themselves.
• Authoring tools conventionally emit at most one FF 03 per track
  at tick 0, but the spec does not constrain count or placement.
  The helper surfaces every occurrence so callers that only want
  the first name per track can collect into a
  HashMap<usize, TrackNameEvent> keyed on TrackNameEvent::track,
  while callers tracking renames over time read the full Vec. On
  a format-0 file the single track's FF 03 is conventionally read
  as the sequence title.
• TrackNameEvent::text_bytes() borrows the raw text payload
  unchanged (the SMF spec leaves the encoding unspecified —
  historically Latin-1, modern DAWs emit UTF-8). text_lossy()
  returns Cow<str> using String::from_utf8_lossy, so invalid
  UTF-8 surfaces as U+FFFD replacement characters rather than
  panicking — convenient default for callers that only need the
  human-readable track label.
• Cost is linear in the total event count and bounded above by the
  parser's existing MAX_EVENTS_PER_FILE cap; the helper does not
  introduce a new allocation ceiling.
• 6 new unit tests in src/smf.rs::tests cover: empty input,
  single name at tick 0, per-track names on a format-1 two-track
  file (Drums / Bass), two FF 03 events on one track in time
  order (Intro at tick 0, Main at tick 480), stable sort at the
  same tick across two tracks, filtering against the full
  text-meta neighbourhood (FF 01 general text, FF 02 copyright,
  FF 04 instrument name, FF 05 lyric, FF 06 marker, FF 07
  cue — with cross-checks that the marker / lyric / cue helpers
  stay uncontaminated), and text_lossy() resilience against
  non-UTF-8 bytes. Brings the in-crate unit suite from 333 to 339
  unit tests, all passing under cargo test.
• Docstring cross-links: SmfFile::lyrics() and
  SmfFile::cue_points() now point at SmfFile::track_names in
  their "distinct from" enumerations, so the doc graph between the
  six text-meta helpers stays bidirectionally connected.

Round 186 — SmfFile::cue_points() iteration helper (FF 07)

• New smf::CueEvent { tick, track, text } plus
  SmfFile::cue_points() -> Vec<CueEvent>. Collects every cue-point
  meta event (FF 07 len text, the Standard MIDI File 1.0
  film-score / theatrical sync convention) from every track, pins
  each one to the absolute tick of its parent track via cumulative
  TrackEvent::delta sums, then merges the per-track sequences
  with a stable sort by tick — track 0 wins over track 1 at the
  same tick, matching the same merge rule used by
  SmfFile::markers() / SmfFile::lyrics() /
  SmfFile::tempo_map() / SmfFile::time_signatures() /
  SmfFile::key_signatures() and by scheduler.rs §"merged event
  list, sorted by absolute tick".
• Only FF 07 is selected. Other text-kind meta events
  (FF 01 general text, FF 02 copyright, FF 03 track name,
  FF 04 instrument name, FF 05 lyric, FF 06 marker, …) are
  filtered out so callers driving external synchronisation (scene
  change, SFX trigger, video cue) don't have to discriminate
  themselves.
• CueEvent::text_bytes() borrows the raw text payload unchanged
  (the SMF spec leaves the encoding unspecified — historically
  Latin-1, modern editors emit UTF-8). text_lossy() returns
  Cow<str> using String::from_utf8_lossy, so invalid UTF-8
  surfaces as U+FFFD replacement characters rather than panicking
  — convenient default for callers that only need the human-readable
  cue name.
• Cost is linear in the total event count and bounded above by
  the parser's existing MAX_EVENTS_PER_FILE cap; the helper does
  not introduce a new allocation ceiling.
• 8 new unit tests in src/smf.rs::tests cover: empty input,
  single cue at tick 0, three-cue in-order sequence (the
  Intro / SceneA / SceneB shape), multi-track merge order,
  stable sort at the same tick, filtering against neighbouring
  text kinds (FF 03 track name, FF 05 lyric, FF 06 marker —
  with a cross-check that the marker and lyric helpers stay
  uncontaminated), absolute-tick accounting through running-status
  channel events, and text_lossy() resilience against non-UTF-8
  bytes. Brings the in-crate suite from 323 to 331 unit tests, all
  passing under cargo test -p oxideav-midi.
• Docstring cross-links: SmfFile::markers() and SmfFile::lyrics()
  now point at SmfFile::cue_points() for the film-score sync
  stream so callers searching either doc find the cue companion.

Round 182 — SmfFile::lyrics() iteration helper (FF 05)

• New smf::LyricEvent { tick, track, text } plus
  SmfFile::lyrics() -> Vec<LyricEvent>. Collects every lyric meta
  event (FF 05 len text, the karaoke .kar syllable convention)
  from every track, pins each one to the absolute tick of its
  parent track via cumulative TrackEvent::delta sums, then merges
  the per-track sequences with a stable sort by tick — track 0
  wins over track 1 at the same tick, matching the same merge rule
  used by SmfFile::markers() / SmfFile::tempo_map() /
  SmfFile::time_signatures() / SmfFile::key_signatures() and
  by scheduler.rs §"merged event list, sorted by absolute tick".
• Only FF 05 is selected. Other text-kind meta events
  (FF 01 general text, FF 02 copyright, FF 03 track name,
  FF 04 instrument name, FF 06 marker, FF 07 cue point, …)
  are filtered out so karaoke callers iterating syllables don't
  have to discriminate themselves.
• LyricEvent::text_bytes() borrows the raw text payload
  unchanged (the SMF spec leaves the encoding unspecified —
  historically Latin-1, modern files emit UTF-8). text_lossy()
  returns Cow<str> using String::from_utf8_lossy, so invalid
  UTF-8 surfaces as U+FFFD replacement characters rather than
  panicking — convenient default for callers that only need the
  human-readable text.
• Cost is linear in the total event count and bounded above by
  the parser's existing MAX_EVENTS_PER_FILE cap; the helper
  does not introduce a new allocation ceiling.
• 8 new unit tests in src/smf.rs::tests cover: empty input,
  single syllable at tick 0, four-syllable in-order sequence
  (the "Twinkle, Twinkle" .kar shape), multi-track merge order,
  stable sort at the same tick, filtering against neighbouring
  text kinds (FF 03 track name, FF 06 marker), absolute-tick
  accounting through running-status channel events, and
  text_lossy() resilience against non-UTF-8 bytes. Brings the
  in-crate suite from 315 to 323 unit tests, all passing under
  cargo test -p oxideav-midi.
• Docstring cross-link: SmfFile::markers() now points at
  SmfFile::lyrics() for the karaoke syllable stream so callers
  searching the marker docs find the lyric companion.

Round 176 — SmfFile::markers() iteration helper (FF 06)

• New smf::MarkerEvent { tick, track, text } plus
  SmfFile::markers() -> Vec<MarkerEvent>. Collects every marker
  meta event (FF 06 len text, the DAW song-section convention)
  from every track, pins each one to the absolute tick of its
  parent track via cumulative TrackEvent::delta sums, then merges
  the per-track sequences with a stable sort by tick — track 0
  wins over track 1 at the same tick, matching the same merge
  rule used by SmfFile::tempo_map() /
  SmfFile::time_signatures() / SmfFile::key_signatures() and
  by scheduler.rs §"merged event list, sorted by absolute tick".
• Only FF 06 is selected. Other text-kind meta events
  (FF 03 track name, FF 05 lyric, FF 07 cue point, …)
  are filtered out so callers iterating section labels don't have
  to discriminate themselves.
• MarkerEvent::text_bytes() borrows the raw text payload
  unchanged (the SMF spec leaves the encoding unspecified —
  historically Latin-1, modern DAWs emit UTF-8). text_lossy()
  returns Cow<str> using String::from_utf8_lossy, so invalid
  UTF-8 surfaces as U+FFFD replacement characters rather than
  panicking — convenient default for callers that only need the
  human-readable label.
• Cost is linear in the total event count and bounded above by
  the parser's existing MAX_EVENTS_PER_FILE cap; the helper
  does not introduce a new allocation ceiling.
• 8 new unit tests in src/smf.rs::tests cover: empty input,
  single marker at tick 0, multiple per-track markers in order,
  multi-track merge order, stable sort at the same tick,
  filtering against neighbouring te...

v0.0.2

Added

• (midi) GM2 Global Parameter Control (Universal Real-Time SysEx 04 05)
• (midi) Master Balance (Universal Real-Time SysEx 04 02)
• (midi) Data Increment / Decrement (CC 96/97) per RP-018

Other

• round 98 — MIDI Tuning Standard (MTS) microtuning
• Round 95: SFZ-side filter envelope + fil_type + cutoff wiring
• rewrite release-envelope comment to remove FluidSynth citation
• EG2 + 2-pole resonant low-pass filter on the shared SamplePlayer (round 91)
• DLS art1/art2 articulation interpretation (round 80)
• round 75 — MPE + RPN 1/2/5 + CA-25 master tuning + master volume SysEx
• registry calls: rename make_decoder/make_encoder → first_decoder/first_encoder

Round 114 — GM2 Global Parameter Control (Universal Real-Time SysEx 04 05)

• New mixer::GmEffects carries the GM2 system-wide Reverb + Chorus
  parameters edited by the Global Parameter Control Universal
  Real-Time SysEx message (F0 7F <dev> 04 05 …, MMA CA-024). The two
  GM2-reserved slots are 0101 (Reverb) and 0102 (Chorus). Each raw
  7-bit parameter value is decoded to its engineering unit with the
  CA-024 "Recommended Practice for Reverb and Chorus Parameters (from
  General MIDI Level 2)" formulas:
  • Reverb pp=0 Type (select), pp=1 Time
    rt = exp((val − 40) · 0.025) s.
  • Chorus pp=0 Type (select), pp=1 Mod Rate mr = val · 0.122 Hz,
    pp=2 Mod Depth md = (val + 1) / 3.2 ms, pp=3 Feedback
    fb = val · 0.763 %, pp=4 Send-to-Reverb ctr = val · 0.787 %.
• New Mixer::set_gm_reverb_param(pp, val) /
  Mixer::set_gm_chorus_param(pp, val) apply one parameter-value pair;
  unrecognised parameter ids are ignored per CA-024 ("only that
  parameter-value pair should be ignored"). Mixer::gm_effects()
  exposes the current state and Mixer::reset_gm_effects() restores
  the GM2 recommended initial defaults (Reverb Type 4 Large Hall,
  Chorus Type 2 Chorus 3, with the per-type table values).
• The scheduler's Universal Real-Time dispatch now routes sub-ID#2
  05 (dispatch_global_parameter_control): it parses the Slot Path
  Length / Parameter-ID Width / Value Width header, requires the
  GM2-reserved slot path (length 1, Slot MSB 1), reads the MSB-first
  parameter ids and LSB-first values across the pair list to EOX, and
  routes each into the reverb/chorus setters by Slot LSB. Non-GM2 slot
  paths (Slot MSB ≠ 1 or length ≠ 1) are ignored.
• GM 1 / GM 2 System On / GM System Off now also reset the GM2 effect
  parameters to their CA-024 defaults.
• Tests: GM2 defaults; reverb type+time decode; all five chorus
  parameter decodes; non-GM2 slot ignored; unknown-parameter-in-a-pair
  ignored (rest applied); GM-on resets effects. (+7 lib tests.)
• The decoded parameters are observable program state; a reverb/chorus
  DSP send is intentionally deferred to a later round.

Round 105 — Master Balance (Universal Real-Time SysEx 04 02)

• New Mixer::set_master_balance_14(value) /
  Mixer::master_balance_14() carry the device-level Master Balance
  scalar from the MIDI 1.0 Detailed Specification v4.2.1 §"DEVICE
  CONTROL — MASTER VOLUME AND MASTER BALANCE" (p.57). The 14-bit
  value is stored verbatim with 0x0000 = hard left,
  0x2000 = centre, 0x3FFF = hard right; the setter clamps inputs
  above 0x3FFF to the spec maximum.
• New Mixer::master_balance_gains() returns the per-side
  multipliers (left, right) that the mix loop folds into every
  voice's gain — (1.0, 1.0) at centre, (1.0, 0.0) at hard left,
  (0.0, 1.0) at hard right, and a linear ramp on the far side
  between centre and each extreme (the near side stays at unity).
  This is the textbook "balance between two sound sources" law M1
  v4.2.1 §"BALANCE" describes for CC 8, applied here as the
  device-level analog.
• The stereo + mono branches of Mixer::mix_stereo now multiply
  every voice by these master-balance gains. The values are hoisted
  out of the per-slot loop, alongside the existing master-volume
  scalar, so the per-voice arithmetic gains a single extra f32
  multiply per side and the default 0x2000 setting produces an
  output buffer byte-identical to the pre-round-105 mix (asserted by
  the new master_balance_centre_matches_pre_balance_output test).
• scheduler::dispatch_universal_sysex recognises 04 02 lsb msb
  in the Universal Real-Time area and forwards the combined 14-bit
  value via set_master_balance_14. GM 1 / GM 2 System On / GM
  System Off resets now also restore Master Balance to centre
  (0x2000), matching the rest of the master-state reset surface.
• 12 new tests (9 mixer, 3 scheduler): default-centre +
  unity-gains, hard-left mutes right, hard-right mutes left,
  half-left/half-right ramp arithmetic, the clamp-above-14-bit
  guard, the per-side zeroing of the mix output at each extreme, the
  centre-equals-default-output regression, and the three scheduler
  SysEx routings (centre / hard left / hard right). The existing
  universal_gm_on_sysex_resets_state test gained an additional
  Master-Balance-set-then-reset assertion.

Round 102 — Data Increment / Data Decrement (RP-018)

• New Mixer::data_inc_dec(channel, step) implements the Data Increment
  (CC 96) / Data Decrement (CC 97) response from the MMA Response to
  Data Inc/Dec Controllers recommended practice
  (docs/audio/midi/recommended-practices/rp18.pdf, RP-018). Per the
  spec the controller's value byte is don't care; the scheduler passes
  a fixed +1 step for CC 96 and -1 for CC 97. Each step adjusts the
  sub-field RP-018 prescribes for the currently-selected RPN:
  • RPN 0 (Pitch Bend Sensitivity): step the LSB (cents). Because
    the mixer stores the combined pitch_bend_range_cents
    (= semitones·100 + cents), ±1 performs the spec's
    "LSB-wraps-into-MSB at 100" carry automatically (RP-018 worked
    example: two CC 96 = +2 cents; 200 → 199 borrows down into 1
    semitone + 99 cents). Clamped to >= 1 so the range never reaches
    zero, and the live pitch bend is re-applied to held voices.
  • RPN 1 (Channel Fine Tuning): step the LSB of the 14-bit
    fine-tune accumulator; the cents view is re-derived and routed to
    held voices.
  • RPN 2 (Channel Coarse Tuning): step the MSB (= one semitone) per
    the 4.2-Addendum rule RP-018 cites, clamped to the CA-25 signed
    range −64..=+63.
  • RPN 5 (Modulation Depth Range, CA-26): step the cents field, the
    RP-018 default for future Registered Parameters, clamped to the
    existing 0..=2400 envelope.
  • RPN Null (0x3FFF) and any unmodelled / NRPN selection are a
    no-op, mirroring set_data_entry's null guard. NRPNs (CC 98/99) are
    not modelled, so a step issued under an NRPN selection does nothing.
• scheduler::dispatch routes CC 96 → data_inc_dec(ch, 1) and CC 97 →
  data_inc_dec(ch, -1).
• 11 new tests (10 mixer, 1 scheduler): RPN-0 cent-step +
  LSB-wraps-into-MSB carry, RPN-0 decrement borrow, value-byte-ignored
  contract, RPN-1 fine-tune LSB step, RPN-2 semitone step + signed-range
  clamp, RPN-5 cent step, RPN-Null no-op, RPN-0 clamp-above-zero, held-
  voice bend re-apply on range widen, and the scheduler CC 96/97 routing
  (with a deliberately nonsense data byte to prove it is ignored).

Round 98 — MIDI Tuning Standard (MTS) microtuning

• New tuning module implements the retuning surface from the MMA
  MIDI Tuning Messages specification
  (docs/audio/midi/extensions/MIDI-Tuning-Updated-Specification.pdf,
  incorporating CA-020 / CA-021 / RP-020). TuningTable holds two
  layers of microtuning state — a global 128-entry key-based table and
  per-channel 12-entry scale/octave tables — both expressed as signed
  cents added to a key's 12-tone-equal-temperament pitch. Defaults to
  equal temperament everywhere, so a synth that never receives an MTS
  message renders bit-identically to the pre-MTS path.
• Data-format decoders, each with worked-example unit tests against the
  spec's tables: freq_word_to_cents_offset (3-byte
  semitone + fraction14/16384 frequency word → cents offset from the
  addressed key, with the reserved 7F 7F 7F "no change" word
  returning None); scale_octave_1byte_to_cents (00 = -64 c,
  40 = 0 c, 7F = +63 c); scale_octave_2byte_to_cents (14-bit,
  0x0000 = -100 c, 0x2000 = 0 c, 0x3FFF = +100 c); and
  scale_octave_channel_mask (the ff gg hh 3-byte channel bitmap,
  with ff bits 2–6 reserved → must not light any channel).
• Mixer carries a TuningTable; the per-key offset is folded into
  every voice-pitch composition site (note-on + the two
  set_pitch_bend re-apply paths). Drum channel (MIDI 10 = index 9) is
  exempt from retuning, matching the existing CA-25 master-tuning
  exemption. New public API: set_key_tuning_word,
  set_scale_octave_tuning, reset_tuning, tuning(). The real-time
  message forms re-apply pitch to sounding voices (live = true); the
  non-real-time "setup" forms update only the stored table.
• scheduler::dispatch_universal_sysex now routes sub-ID#1 08 (MIDI
  Tuning Standard) in both the Universal Real-Time (7F) and
  Non-Real-Time (7E) areas: Single-Note Tuning Change (sub-ID#2
  02) and its bank form (07), and Scale/Octave Tuning 1-byte
  (08) and 2-byte (09) forms. Multi-change single-note messages
  (ll entries) and truncated/over-promised buffers are bounds-checked
  so malformed input cannot read past the payload. GM 1 / GM 2 System
  On / GM System Off now also reset MTS tuning to equal temperament.
• 25 new tests (12 tuning unit, 7 mixer, 6 scheduler) covering
  the decoders, table summation, live vs. setup re-apply, drum-channel
  exemption, pitch-bend summation, channel-mask selection, GM reset,
  and truncated-message safety.

Round 95 — SFZ-side filter envelope + fil_type + cutoff wiring

• New FilterType enum on instruments::sample_voice covers the six
  SFZ v1 fil_type values documented in
  `docs/a...

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 r...

v0.0.0

chore: Release package oxideav-midi version 0.0.0