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/audio/midi/instrument-formats/sfz-legacy.html "Filter type"
  table: lpf_1p / hpf_1p (one-pole, 6 dB/oct, resonance ignored)
  and lpf_2p / hpf_2p / bpf_2p / brf_2p (two-pole, 12 dB/oct).
  Default is TwoPoleLowPass, which preserves round-91 SF2 / DLS
  behaviour exactly. FilterType::parse_sfz() honours the
  case-insensitive opcode string + falls back to lpf_2p on unknown
  values per the SFZ default convention.

• FilterParams gains a kind: FilterType field. SF2 / DLS construct
  the struct without overriding it (they inherit TwoPoleLowPass); the
  instruments::articulation::Articulation::filter() helper also pins
  the kind explicitly to make the SF2/DLS commitment to a single shape
  permanent against any future Default flip.

• SamplePlayer::update_filter_coeffs switches on filter_kind. The
  one-pole arms compute tan(ω/2)-prewarped bilinear coefficients with
  b2 = a2 = 0 so the same direct-form-1 tick() math still applies
  (no per-sample branch). The 2-pole arms keep the round-91 RBJ-cookbook
  denominators (a0 = 1 + α, a1 = -2·cos(ω), a2 = 1 - α) and switch
  numerators per shape: low-pass (1-cos)/2, 1-cos, (1-cos)/2,
  high-pass (1+cos)/2, -(1+cos), (1+cos)/2, band-pass
  sin/2, 0, -sin/2, band-reject 1, -2cos, 1. All derivations are
  the project's own RBJ math (bilinear transform of the analog
  prototypes) — SF2 §8.1.3 + §9.7 + SFZ-legacy fil_type row only
  specify the response shape and slope, not the discrete realisation.

• SamplePlayer::new's needs_filter gate now respects non-low-pass
  shapes: a fil_type of hpf_* / bpf_2p / brf_2p is not "open"
  at the SF2 13500-cents sentinel (it would still attenuate the audible
  band), so the biquad allocation also fires whenever
  cfg.filter.kind is anything other than the two low-pass variants.

• instruments::sfz::build_filter_from_opcodes() parses cutoff=
  (Hz, with a cents = 1200·log2(fc_hz / 8.176) bridge into SF2
  absolute cents — clamped into the §8.1.3 useful range 1500..=13500),
  resonance= (dB → centibels at cb = dB · 10, clamped to the
  SFZ-spec range 0..=40 dB), and fil_type= (via
  FilterType::parse_sfz). Aliases (cutoff2, resonance2,
  filtype) are honoured per the opcode index.

• instruments::sfz::build_mod_env_from_opcodes() parses the SFZ v1
  fileg_* Envelope-Generator opcodes (fileg_delay /
  fileg_attack / fileg_hold / fileg_decay / fileg_sustain /
  fileg_release + fileg_depth for the routing depth) and their
  SFZ v2 fil_* aliases (fil_delay etc.). fileg_sustain is in
  percent and maps to the SamplePlayer's 0..=1 fraction;
  fileg_depth is clamped into the documented -12000..=12000
  range and dropped into ModEnvParams::to_filter_cents.

• instruments::sfz::build_config_for_region calls the two helpers
  instead of plumbing Default::default() for mod_env / filter.
  Regions without fileg_* / fil_type / cutoff / resonance
  opcodes still render bit-identically to the round-91 path — the
  SF2 "filter open" sentinel (13500 cents) + the inert ModEnvParams
  default fall straight through build_filter_from_opcodes() /
  build_mod_env_from_opcodes().

• 18 new tests in the crate:

  • instruments::sample_voice::tests::high_pass_attenuates_below_cutoff
  • instruments::sample_voice::tests::band_pass_peaks_at_cutoff
  • instruments::sample_voice::tests::band_reject_kills_signal_at_cutoff
  • instruments::sample_voice::tests::one_pole_low_pass_attenuates_high_frequencies
  • instruments::sfz::tests::hz_to_filter_cents_round_trips_known_values
  • instruments::sfz::tests::build_filter_from_opcodes_honours_cutoff_resonance_filtype
  • instruments::sfz::tests::build_filter_from_opcodes_defaults_to_open_lpf_2p
  • instruments::sfz::tests::build_filter_from_opcodes_clamps_resonance_into_range
  • instruments::sfz::tests::filter_type_parse_covers_all_sfz_v1_values
  • instruments::sfz::tests::build_mod_env_from_opcodes_honours_fileg_set
  • instruments::sfz::tests::build_mod_env_from_opcodes_aliases_v2_names
  • instruments::sfz::tests::build_mod_env_default_is_inert
  • instruments::sfz::tests::build_mod_env_clamps_depth
  • instruments::sfz::tests::full_template_smoke_drops_cutoff_into_sample_player
  • instruments::sfz::tests::sfz_with_low_cutoff_attenuates_high_frequencies
  • instruments::sfz::tests::sfz_high_pass_inverts_attenuation
  • instruments::sfz::tests::sfz_fileg_sweeps_cutoff_open_during_attack
  • instruments::sfz::tests::build_config_for_region_default_filter_open

  All 228 lib tests + 14 integration tests pass (242 total, 0 ignored).

• Sources: docs/audio/midi/instrument-formats/sfz-opcodes-index.html
  (Aria opcode reference: cutoff, resonance, fil_type, fileg_*

  • fil_* aliases with documented defaults / ranges / SFZ-version
    tags), docs/audio/midi/instrument-formats/sfz-legacy.html
    ("Filter type" table enumerating the six fil_type values + the
    "filter disabled" semantics for an absent cutoff opcode),
    docs/audio/midi/instrument-formats/sf2-spec-2.04.pdf §8.1.3
    (initialFilterFc / initialFilterQ unit conventions reused
    unchanged from round 91).

Round 91 — EG2 + 2-pole resonant low-pass filter on the shared SamplePlayer

• New ModEnvParams + FilterParams structs on
  instruments::sample_voice. ModEnvParams mirrors the existing
  EnvelopeParams DAHDSR shape but treats sustain_level as the
  post-decrease linear fraction per SF2 v2.04 §8.1.3 sustainModEnv
  ("decrease in level expressed in 0.1 % units", with 0 = peak and 1000
  = silence) and carries to_filter_cents — the per-mod-env-unit
  contribution to the filter cutoff per SF2 gen 11 modEnvToFilterFc.
  FilterParams carries cutoff_cents (SF2 absolute cents re. 8.176
  Hz; default 13500 = "filter open" sentinel from §8.1.3 gen 8) and
  q_centibels (0 = Butterworth, per §8.1.3 gen 9).
• SamplePlayer::new allocates a per-voice 2-pole resonant low-pass
  biquad only when the bank actually wants the filter (cutoff <
  13000 cents or mod-env-to-filter routing depth > 200 cents).
  Otherwise the per-sample biquad work is skipped entirely so
  unconfigured SFZ / DLS regions render bit-identically to the
  round-80 pre-filter path.
• Biquad coefficients computed from the SF2 §8.1.3 cents reference
  fc_hz = 8.176 * 2^(cents/1200) (clamped to the spec useful range
  1500..=13500, then clipped at 0.99 * Nyquist) and the RBJ
  cookbook low-pass formulas: a0 = 1 + α, a1 = -2·cos(ω),
  a2 = 1 - α, b0 = b2 = (1 - cos(ω))/2, b1 = 1 - cos(ω), where
  α = sin(ω)/(2Q). Q (centibels) → linear q_lin = √(½) · 10^(cb/200)
  with a 0.1..=16.0 clamp so runaway resonance can't produce NaN
  coefficients. The RBJ derivation is the project's own clean-room
  math (bilinear transform of the analog 2-pole low-pass
  H(s) = ω₀² / (s² + (ω₀/Q)s + ω₀²)) — SF2 §8.1.3 explicitly
  leaves the filter implementation to the renderer per §9.7.
• SamplePlayer::render evaluates the mod-env DAHDSR each sample,
  adds mod_env_level * mod_env_to_filter_cents to the initial
  cutoff, and lazily recomputes biquad coefficients only when the
  live cutoff drifts > 50 cents from the last computed value (cheap
  perceptual gate that keeps the inner loop multiplication-only
  during the delay / hold / sustain phases when the mod-env is
  constant). The filter sits between sample fetch and amplitude
  envelope, so EG1 amplitude shaping is post-filter as expected.
• SamplePlayer::release now captures the EG2 release-start level
  separately from the EG1 release-start level so the filter cutoff
  tails off from wherever the mod-env was at note-off, not from
  peak — a discontinuity-free release for both amplitude and filter.
• New Articulation::mod_env() -> ModEnvParams +
  Articulation::filter() -> FilterParams helpers on
  instruments::articulation. mod_env() converts the DLS EG2
  destinations (CONN_DST_EG2_DELAYTIME / _ATTACKTIME /
  _HOLDTIME / _DECAYTIME / _SUSTAINLEVEL / _RELEASETIME
  plus the SRC_EG2 → DST_FILTER_CUTOFF routing depth) into the
  SamplePlayer-side ModEnvParams; filter() maps
  CONN_DST_FILTER_CUTOFF / CONN_DST_FILTER_Q into FilterParams,
  falling back to the SF2 "filter open" sentinel (13500 cents) when
  the region carries no filter blocks.
• instruments::dls::build_dls_config plumbs art.mod_env() +
  art.filter() into the SamplePlayerConfig. A DLS bank with no
  art2 filter blocks produces the same audio it did pre-round-91;
  a bank with an art2 SRC_EG2 → DST_FILTER_CUTOFF block sweeps
  the cutoff exactly as the spec describes.
• instruments::sfz::build_config_for_region populates the new
  fields with Default::default() — SFZ fileg_* / fil_type /
  cutoff opcodes are not yet plumbed; SFZ banks render
  bit-identically to the round-80 path until a future round wires
  the SFZ filter opcodes.
• 12 new tests across the crate:
  • instruments::sample_voice::tests::default_filter_is_inert_no_biquad_allocated
  • instruments::sample_voice::tests::low_cutoff_filter_attenuates_high_frequencies
  • instruments::sample_voice::tests::mod_env_dahdsr_shape_matches_spec
  • instruments::sample_voice::tests::eg2_filter_sweep_changes_spectrum_over_note
  • instruments::sample_voice::tests::high_q_filter_resonates_at_cutoff
  • instruments::sample_voice::tests::release_captures_mod_env_level
  • instruments::articulation::tests::default_mod_env_is_inert
  • instruments::articulation::tests::default_filter_is_open_sentinel
  • instruments::articulation::tests::eg2_to_filter_routing_lands_on_mod_env
  • instruments::articulation::tests::filter_cutoff_override_lands_on_filter
  • instruments::articulation::tests::eg2_attack_time_lands_on_mod_env_attack
  • tests::dls_articulation::dls_articulation_eg2_sweeps_filter_cutoff_open
    (integration: builds a DLS bank with SRC_EG2 → DST_FILTER_CUTOFF
    • slow EG2 attack, renders through the full MidiDecoder, asserts
      the late-window RMS is > 1.6× the early-window RMS).
• Test count moves from 212 → 224.
• Spec backing: SF2 v2.04 §8.1.3 generators 8 + 9 + 11 + 25–30
  (docs/audio/midi/instrument-formats/sf2-spec-2.04.pdf); DLS Level
  2.2 v1.0 Amendment 2 §1.5.2 + Tables 5–6 + 8–10 + 1.13 + 1.14
  (docs/audio/midi/instrument-formats/dls2amd2(all)a(pub).pdf); DLS
  Level 1 v1.1b "Device Architecture" + "Articulation"
  (docs/audio/midi/instrument-formats/dls1v11b.pdf).

Round 80 — DLS art1 / art2 articulation interpretation

• New module instruments::articulation interpreting DLS Level 1 and
  Level 2 connection blocks at voice-build time. Backed by MMA DLS
  Level 1 v1.1b (docs/audio/midi/instrument-formats/dls1v11b.pdf,
  Table 1 + 2 of the Device Architecture section) and MMA DLS Level 2.2
  v1.0 Amendment 2 (docs/audio/midi/instrument-formats/dls2amd2(all)a(pub).pdf,
  Tables 5–10).
• Named constants for every CONN_SRC_* / CONN_DST_* / CONN_TRN_*
  enum from DLS2 Tables 8 + 9 + 10, plus ABSOLUTE_ZERO sentinel,
  exported from instruments::articulation so callers can build
  blocks programmatically (used by the new
  crates/oxideav-midi/tests/dls_articulation.rs integration test).
• New Articulation::evaluate(region_blocks, instrument_blocks) -> Articulation walks the region-level then instrument-level
  connection lists, overlaying every recognised connection on top of
  the spec defaults (DLS2 Tables 5 + 6). Per the DLS spec, a region
  block overrides the corresponding instrument-level block.
• Supported SRC_NONE → DST_x connections (the "absolute default
  override" branch): Vol EG (EG1) delay / attack / hold / decay /
  sustain / release; Mod EG (EG2) delay / attack / hold / decay /
  sustain / release (raw, surfaced for a later round); modulator LFO
  frequency + start delay; vibrato LFO frequency + start delay;
  filter cutoff (DST_FILTER_CUTOFF) and Q; tuning (DST_PITCH);
  per-region gain (DST_GAIN); pan (DST_PAN).
• Supported modulator routings: SRC_LFO → DST_PITCH (vibrato depth
  on DLS1-style banks where LFO and vibrato share a source);
  SRC_LFO → DST_GAIN (tremolo depth); SRC_VIBRATO → DST_PITCH
  (dedicated DLS2 vibrato depth — wins over SRC_LFO → DST_PITCH
  when both are present); SRC_EG2 → DST_PITCH and SRC_EG2 → DST_FILTER_CUTOFF (mod-env routings, raw); SRC_KEYONVELOCITY → DST_EG1_ATTACKTIME (velocity-dependent attack, raw).
• DlsInstrument::make_voice now calls
  Articulation::evaluate(&region.articulation, &inst.articulation)
  and folds the resulting [EnvelopeParams] + [VibratoParams] +
  tuning cents + gain multiplier into the SamplePlayerConfig. An
  empty lart list still falls back to the SamplePlayer defaults so
  banks with no articulation are bit-identical to round-75 output.
• Unit conversions: time-cents → seconds (DLS §1.14.3, clamped at
  60 s), absolute-pitch → cents (DLS §1.14.1, clamped at ±14 400),
  absolute-pitch → Hz (LFO frequency, clamped at 50 Hz), gain → linear
  amplitude (DLS §1.14.4, clamped at -96..+48 dB), sustain-percent
  → 0..=1, pan-percent → ±50.
• New integration test tests/dls_articulation.rs exercises the full
  SMF → scheduler → DLS bank → articulation → SamplePlayer → PCM path:
  a SRC_NONE → DST_PITCH @ +1200 cents block produces an audibly
  different rendering than the same bank with no lart list (avg
  pointwise PCM diff > 50 LSB); a SRC_NONE → DST_EG1_RELEASETIME
  block at +2 s sustains a louder release tail than the default 100 ms.
• 9 new unit tests in instruments::articulation: spec-default
  fallback when both lists are empty; region overrides instrument;
  instrument-level fallback when region is silent; tuning cents
  conversion; LFO → pitch routes to the vibrato depth; DLS2 vibrato
  source takes precedence over the mod LFO source; gain destination
  attenuates correctly (-6 dB → 0.5012 linear); ABSOLUTE_ZERO sentinel
  is skipped without overriding the default; unrecognised connections
  are dropped silently.

Total test count: 199 lib + 13 integration = 212 (up from 190 + 11 in
round 75).

Round 75 — MPE + RPN expansion + Master Tuning / Master Volume SysEx

• MIDI Polyphonic Expression (MPE) v1.1 end-to-end
  (docs/audio/midi/extensions/M1-100-UM_v1-1_MIDI_Polyphonic_Expression_Specification.pdf):
  the MPE Configuration Message (RPN 0x0006 on channel 0 = Lower
  Manager / channel 15 = Upper Manager) configures one or two zones
  via [Mixer::set_mpe_zone]. Per §2.2.5 the receiver sets Manager
  Channel PB Sensitivity to 2 semitones and every Member Channel to
  48 semitones at MCM time. Per §2.2.7 Polyphonic Key Pressure on a
  Member is silently dropped (the spec says it "shall not be sent").
  Per Appendix C, Member Channel Pitch Bend sums in cents with the
  Manager Channel's bend before reaching the held voice. Per §2.2.3,
  a zone reconfiguration stops every Sounding Note on the affected
  channels and resets their controllers. New types
  [Mixer::MpeZone] / [Mixer::MpeRole] / [Mixer::MpeZoneKind]
  surface the zone topology to callers.
• CC #74 ("third dimension of control") routed through the new
  [Voice::set_timbre] hook. MPE Manager broadcasts to every voice
  in the zone; Member only reaches its own voice. Non-MPE channels
  route plainly.
• RPN 1 (Channel Fine Tuning) — 14-bit data-entry value maps
  linearly to ±100 cents (centre 0x40/0x00). Stored as both the raw
  accumulator and the derived cents view on ChannelState so an
  MSB-then-LSB sequence composes bit-exact.
• RPN 2 (Channel Coarse Tuning) — CC 6 MSB sets signed semitone
  offset centred on 0x40 (-64..=+63). CC 38 LSB ignored per spec
  ("the LSB is always 0").
• RPN 5 (Modulation Depth Range) per
  docs/audio/midi/recommended-practices/ca26-RPN05-Modulation-Depth-Range.pdf:
  CC 6 sets whole-cent range, CC 38 sets fractional cents (0..=99).
  Default 50 cents matches the GM 2 recommended practice. Clamped
  to ≤ 2400 cents (±2 octaves) so a stray CC 6 = 127 can't pop the
  timbre out of audibility.
• CC 1 (Modulation Wheel) routed via
  [Mixer::set_mod_wheel] → [Voice::set_mod_depth_cents] using
  the channel's RPN-5 range. Applied to every held voice plus
  picked up at note-on for new voices.
• Universal Real-Time SysEx — Master Volume (F0 7F <dev> 04 01 lsb msb F7): the 14-bit value applies as a multiplicative global
  gain at mix time.
• Universal Real-Time SysEx — Master Fine Tuning / Master Coarse
  Tuning per
  docs/audio/midi/recommended-practices/ca25-Master-Fine-Coarse-Tuning-SysEx-Message.pdf:
  Fine is ±100 cents centred on 0x40/0x00 (formula
  100/8192 × (value - 0x2000)); Coarse is a signed semitone count
  centred on 0x40 with the LSB always 0 per spec. Both sum with the
  per-channel RPN-1 / RPN-2 tuning + the live pitch bend into the
  effective cents pushed to each voice. Drum channel (MIDI 10 =
  index 9) is exempt from all tuning per CA-25's "MUST NOT result in
  MIDI note-shifting" clause.
• Universal Non-Real-Time SysEx — GM 1 / GM 2 System On + GM
  System Off (sub-IDs 09 01 / 09 02 / 09 03) reset the
  mixer's master state to GM defaults: master volume → max, master
  fine / coarse tuning → centre, all sounding notes off.
• Voice trait extension: set_mod_depth_cents + set_timbre
  default-no-op methods so existing voices keep working unchanged.
• Per-channel state expansion: ChannelState gains
  mod_wheel, mod_depth_range_cents, channel_fine_tune_cents,
  channel_fine_tune_raw_14, channel_coarse_tune_semitones, and
  mpe_role fields. All zero / centre by default so existing tests
  see unchanged routing.
• Tests added: 38 new — 27 lib-side mixer (RPN 1 / 2 / 5
  data-entry, channel + master fine/coarse routing, drum channel
  exemption, master volume scaling, mod-wheel routing, CC 74
  routing, MPE zone assignment + role tagging + PB sensitivity
  defaults + zero-members deactivate + Upper zone, Member + Manager
  PB combining, Manager CC 74 broadcast, Member CC 74 isolation,
  Polyphonic Key Pressure dropped on Member, Member + Manager
  pressure combining, zone-conflict resolution, MCM via the data-
  entry pathway, MCM on a non-Manager channel ignored), 6 lib-side
  scheduler (CC 1 / CC 74 / Master Volume / Master Fine / Master
  Coarse / GM-on routing + MCM via SMF), and 5 new integration
  (tests/mpe_and_master_tuning.rs) checking the public
  MidiDecoder surface. Total: 189 lib + 11 integration = 200
  passing (was 156 + 6 = 162).

Wall respected: every change in this round used only
docs/audio/midi/extensions/M1-100-UM_v1-1_MIDI_Polyphonic_Expression_Specification.pdf
(pages 1–28), docs/audio/midi/recommended-practices/ca25-Master-Fine-Coarse-Tuning-SysEx-Message.pdf,
docs/audio/midi/recommended-practices/ca26-RPN05-Modulation-Depth-Range.pdf,
docs/audio/midi/midi-1.0/Universal-System-Exclusive-Messages.pdf,
plus the in-tree crate sources + oxideav-core's public API. No
external library source consulted, paraphrased, or cross-checked.