MS-20 sound: K35/OTA filters, patch panel, and the Squelch 33 acid synth

CHANGELOG.md

@@ -6,6 +6,31 @@ GitHub are its section here (`tools/release.sh` copies them), so add to
 
 ## Unreleased
 
+- **The SM-24 Mono has a patch panel.** Its new PATCH page holds four
+  cords. Drag a jack from the dock onto a ringed knob (LPF CUT, HPF CUT,
+  PW, VCO2 PITCH, LVL, MG FREQ), or pick SRC and DEST on a cord row. The
+  jacks are FENV, AENV and AENV inverted, MG, MG PUL, S&H, NOISE, KBD, VEL,
+  PRESS, VCO1, VCO2 and VCA OUT. The cords run at audio rate, so VCO2 can
+  FM the cutoff and the output can feed back into the filter. A patch
+  with no cords sounds exactly as before. Five new patches show it off
+  (`patch-*`), with the demo `songs/ms20_patch_bay.py`.
+
+- **Squelch 33, an acid bassline synth.** A new mono machine built on the
+  HOT/WET filter the SM-24 Mono used to have: one SAW or SQR oscillator,
+  CUT, RES, ENV MOD, DECAY, ACCENT and SLIDE. Loud notes (the `!` in a
+  slabkit seq) are accented: louder, brighter, and a run of them climbs
+  higher with RES up. Overlapping notes slide. Presets: `classic-squelch`,
+  `square-bounce`, `wet-scream`, `round-line`; demo `songs/squelch_acid.py`.
+
+- **SM-24 Mono sounds like an MS-20.** Its LPF MODE gains **K35** and
+  **OTA**, the early and late MS-20 filters. Bass stays full at any PEAK,
+  the peak sits on CUT, and both the HPF and LPF resonate up to
+  self-oscillation. K35 is the default: grittier and lopsided when driven.
+  OTA is cleaner and sings. HOT and WET are still there for the acid
+  sound. New presets: `k35-scream`, `k35-round-bass`, `ota-singing-lead`.
+  A new SM-24 starts from a re-voiced init patch; saved songs and presets
+  keep their mode and sound as before.
+
 - **A harsh effects pack** (docs/43).
   - **Octaver** gives one and two octaves down the way analog divider
     pedals do, with a GLITCH knob to make it chatter, plus an

bench/golden/acid.txt

@@ -0,0 +1,5 @@
+# bench goldens for acid: case sha256(f32 L ++ f32 R) [nonharm_db thd_db]
+notes 8bc59d45b766ad9007c4db1dbfcdd24b75f1c631ccc33814cc4ea2cc68ca9f61
+velocity b52ab388bda9c4e65b9ebf3b22a557d2771ad44628677afb3fd1ac0e190fccfd
+held 35a9aa5abc0a2babbc8df5521f5f2bd9614e162a47bd05b034765f7a33802a9c -76.1 5.5
+high 0cd64ed21d5ec2f2fb40eb5b4fbcf3d60c2435d8efaf6f3f7a79a694ac46c5b1 -48.3 1.5

bench/golden/ms20.txt

@@ -1,5 +1,5 @@
 # bench goldens for ms20: case sha256(f32 L ++ f32 R) [nonharm_db thd_db]
-notes a07508ad06ee8723008c3675e0ba735e56105fb80a75298e29acc4df1704d188
-velocity ad9081885f858b343993a2bae2a8933f8152e0db0c9e95a3136cf678511c3859
-held 78fe628e416bc871081d2d1aaaeb60b0df256d812f9314579f7cddebdcfd6ae8 -23.3 -0.8
-high df662da24d2dd7cf90588facedc1e490e6250e35e1621b48657b76468bce0c68 -12.2 0.2
+notes cb771ada7c0d5dde3546729c14c16efa53ee4191994848fccd503cdbe36c67e3
+velocity 8e0cde2f8a794f10a1329f8f17f2ce2742c182977515a8e6ca3dda7acf5fd062
+held 90baac0913232f60198a2b81cae395d4d31eb7cb17e4620418b6cd5026087a15 -50.0 -5.2
+high 0de9e0cb902d5539ae2eff6c17697f52e0c239fc3122aab1a3a93a5db79f56d3 -30.1 -13.7

docs/14-ms20-machine.md

@@ -10,6 +10,72 @@ Scope call: **no full semi-modular patch bay.** We implement the MS-20's
 *hardwired* routing (MG/EG→pitch, MG→cutoff, EG→cutoff/VCA, PWM, ring) with
 fixed intensity controls. A general routing matrix can come later.
 
+## K35 and OTA: the MS-20's own filters (2026-10-21)
+
+HOT and WET sound acid: a TB-303, not an MS-20. Their loop subtracts a
+clipped, 18 Hz high-passed copy of the lowpass from the input. That is the
+ladder structure (the 303 even has the high-pass in its loop). Measured
+with a numpy port of `ms20_lpf.fy` at CUT 500 Hz and a quiet input:
+
+- PEAK costs 16 dB of bass at its old default (0.7), and 22–25 dB at the
+  acid presets.
+- The peak sits 1.5–2.2 octaves above CUT.
+- The peak moves about 0.7 octave with input level, which is the rubbery
+  squelch.
+- The slope is −8 to −11 dB/oct.
+
+Stinchcombe's study of the MS-10/MS-20 filters
+([MS20_study.pdf](http://www.timstinchcombe.co.uk/synth/MS20_study.pdf))
+gives both hardware generations LP = 1/(s² + (a − k)s + 1) and
+HP = s²/(same). Here a = 2⅓ for the Korg35 Sallen-Key and 2 for the
+LM13600 OTA pair. So the DC gain doesn't depend on PEAK, the cutoff
+doesn't move, and the slope is 12 dB/oct.
+`kernels/04-filters/ms20_sk.fy` implements it as **MODE K35** and
+**MODE OTA**. HOT and WET stay as they were.
+
+```
+v      = x − (a + g)·s1 − s2
+bp-lin = (g·h·v + s1) / (1 − g·h·k)          h = 1/(1 + a·g + g²)
+fb     = vd·tanh(k·bp-lin / vd)              feedback diodes
+u      = vf·(tanh((drive·x + fb)/vf + fo) − tanh fo)
+TPT SVF with base damping a → hp, bp, lp
+```
+
+- **K35:** the diodes sit in the forward amplifier (vf 0.55), so loud
+  input distorts every frequency. A small offset (fo 0.12) makes the
+  resonance lopsided: the self-oscillation's 2nd harmonic is −37 to
+  −46 dB. Both are Stinchcombe §4–5.
+- **OTA:** the feedback diodes set the peak and the forward stage is soft
+  (vf 1.6), so the passband stays clean at normal DRIVE. The
+  self-oscillation is symmetric (2nd harmonic below −115 dB).
+- **Both filters resonate.** In these modes the HPF is the same core with
+  its own PEAK. The series HPF → LPF pair runs at 4× between `up4` and
+  `dec4`, and both cutoffs stop at 15 kHz (Korg's spec).
+- **PEAK taper:** k = a·(1 − u²) with u = 1 − (PEAK/2)/0.9, rising
+  linearly past 90%. This is a design choice, not a measured pot. A linear
+  pot repeats the 2026-09-26 OTA probe's "resonance invisible until the
+  top" (+1 dB at noon). This taper gives +6 dB at noon, and
+  self-oscillation from 1.82 of 2.
+
+Measured through the voice in the bench, with noise in (CUT 500 Hz, HPF
+1 kHz):
+
+| | K35 | OTA |
+|---|---|---|
+| bass at 60 Hz, PEAK 0 → 2, small signal | −0.2 … +0.4 dB | −0.1 … +0.2 dB |
+| peak position, PEAK ≥ 1.4 | within 0.02 oct of CUT | within 0.02 oct |
+| slope, 2 → 3 octaves past the corner | −10.7 … −11.8 dB/oct | −11.2 … −12.1 dB/oct |
+| peak at PEAK 1 (noon) | +6.4 dB | +8.1 dB |
+| peak, quiet → loud input, PEAK 1.75 | 23.5 → 8.3 dB | 32.5 → 15.3 dB |
+| HPF peak position / slope | within 0.04 oct / +11.5 … +12.6 | within 0.04 oct / +11.3 … +12.7 |
+| default notes, non-harmonic energy | −31 … −41 dB (HOT −18 … −22) | −31 … −42 dB |
+| cost, ReleaseFast, held note | ~2× HOT (≈ 350 vs 184 ns/smp) | same |
+
+K35 is the default for new instances. The init patch was re-voiced
+because CUT now means the corner: 1200 Hz, PEAK 1, ENV 2 octaves,
+filter decay 0.3 s, sustain 0.2. The acid pair is kept, both here and in
+`ms20_lpf.fy` for a future acid machine.
+
 ## LPF v3: the June probe's driven SVF, in fy (2026-09-27)
 
 The reference is `scratch/ms20_saw_filter_sweeps_f-hot.wav`, rendered by
@@ -285,3 +351,67 @@ hatch) is a good early proof.
 7. Panel: declare the strips (docs/15); custom ADSR cell.
 8. Global spilling (durable follow-up) to optionally collapse hot stages back
    inline; polyphony via a higher-order voice-pool machine (docs/15).
+
+## Squelch 33: the acid filter's own synth (2026-10-22)
+
+HOT/WET stayed in the SM-24 Mono, and the same kernel (`ms20_lpf.fy`,
+unchanged) now also drives `machines/acid`, a 303-style mono voice in
+`kernels/06-voices/acid_voice.fy`: one VCO (saw, or a square with a
+little of the ramp on its tops), the LPF at 4x, a decay-only MEG with
+ENV MOD (5 octaves at full), a VCA that decays over 3 s and closes in
+8 ms on key-up.
+
+- **Accent** is velocity over 0.8. An accented note runs a second MEG
+  at the 200 ms minimum decay; that envelope pushes the VCA directly
+  (x[1 + 1.2 ACC]) and feeds the accent RC (15 ms, 60 ms at full RES),
+  whose output adds 4 octaves x level to the cutoff. With RES up the RC
+  does not empty between notes, so accents in a row stack.
+- **Slide**: a note that overlaps the last (`ctx.legato`) glides on an
+  RC covering 3 time constants in SLIDE, and retriggers nothing.
+
+Measured (slabkit renders, A1 lines, 120-128 BPM):
+
+| | target | result |
+|---|---|---|
+| accent level | +4..6 dB | +4.3..5.3 dB at RES 0.3, +3.3..4.6 at 0.9 |
+| accent attack centroid | brighter | +40% at RES 0.3, +20% at 0.9 |
+| accent stacking, 3 in a row | climbs at high RES | +15% at RES 0.9, +5% at 0.3 |
+| fundamental loss, RES 0 -> 1 | the acid thinning | -13 dB HOT, -13.4 dB WET |
+| loudest partial, CUT 400 | 1.5-2 oct up | 654 Hz HOT, 785 Hz WET |
+| slide 60 ms | lands in ~60 ms | ~50-60 ms, no retrigger |
+| preset loudness on a line | -20 dB RMS +-1 | -20.0 .. -19.6 |
+| cost | | 233 ns/sample ReleaseFast |
+
+## The patch panel (2026-10-22)
+
+Four cords, evaluated every sample in the voice (`ms20-bay`,
+`ms20-cord-to` in `kernels/06-voices/ms20_voice_probe.fy`), declared to
+the panel the way Concoction's matrix is (docs/15 §Modulation): controls
+`ms20-c1-src` / `-dst` / `-amt` ... `ms20-c4-*`, `mod-source` jacks in the
+dock, `mod-dest` rings, the normalled routes (ENV and MG to LPF CUT, EG
+and MG to PITCH, MG to PW) as fixed rows, a PATCH page with the dock,
+the CORDS matrix display and a scope.
+
+- **Jacks** are an array in the voice state (`jk`), one per SRC option,
+  stored each sample and read by a runtime index, which sees the state
+  as the word found it: every source arrives one sample late. That is
+  what lets VCA OUT go back into the filter without a zero-delay loop.
+- **Destinations** and their range at AMT 1: PITCH and VCO2 +-2
+  octaves, PW +-0.45, HPF CUT and LPF CUT +-5 octaves (KBD at AMT 0.2
+  tracks the keyboard 1:1), VCA gain +-1, MG FREQ +-4 octaves, FLT IN the
+  source as audio at the mixer's level, ahead of both filters.
+- **Free when empty.** The bay sits behind `bay-on` (a cord with SRC and
+  DEST set), the per-sample HPF coefficients behind `bay-hpf`; both are
+  versioned `ifte`s, and every unpatched path adds exact zeros, so all
+  goldens stay bit-exact. VCO1's wave became a select so the voice fits
+  fy's 32 versioned bodies (VCO2 wave 4 x MODE 2 x 2 x 2); past that the
+  bay if-converted and cost 609 ns/sample unpatched.
+- Cost, ReleaseFast: 413 ns/sample unpatched, 470 with a cord.
+
+Measured on a held C3 (K35, CUT 600, PEAK 1, ENV 0): MG -> LPF CUT 0.3
+triples the centroid's swing (66 -> 190 Hz); MG -> PITCH 0.05 gives
+vibrato; AENV -> VCA -0.5 takes 6 dB off; VCO2 a fifth up -> LPF CUT
+0.3 doubles the centroid (418 -> 866 Hz) at audio rate; VCA OUT -> FLT IN
+0.8 stays bounded (peak -6.9 dBFS). The `patch-*` presets sit at -21 to
+-22 dB RMS on a bassline (the vowel patch, a band-pass, -22 two octaves
+up).

docs/20-machine-reference.md

@@ -8,6 +8,7 @@ project or preset stores (docs/19): real units for knobs, the option
 
 | id | name | kind | voices | prefix |
 |---|---|---|---|---|
+| [`acid`](#acid) | Squelch 33 | instrument | mono | `—` |
 | [`bass`](#bass) | Fretwire | instrument | 4 | `bs-` |
 | [`concoction`](#concoction) | Concoction | instrument | 8 | `cn-` |
 | [`cream`](#cream) | Mog Passenger | instrument | mono | `cr-` |
@@ -51,6 +52,31 @@ project or preset stores (docs/19): real units for knobs, the option
 | [`verb2`](#verb2) | Verb | effect | — | `verb-` |
 | [`wob2`](#wob2) | Wobble | effect | — | `wob-` |
 
+## acid
+
+**Squelch 33**, instrument, mono.
+
+| param | section | label | range | default | notes |
+|---|---|---|---|---|---|
+| `tune` | VCO | TUNE | -12 … 12 | 0 | linear |
+| `wave` | VCO | WAVE | 0 SAW / 1 SQR | 0 | switch |
+| `cutoff` | VCF | CUT | 30 … 8000 | 260 | exp |
+| `resonance` | VCF | RES | 0 … 1 | 0.6 | linear |
+| `env-mod` | VCF | ENV | 0 … 1 | 0.5 | linear |
+| `decay` | VCF | DEC | 0.2 … 2 | 0.5 | exp |
+| `accent` | PLAY | ACC | 0 … 1 | 0.7 | linear |
+| `slide` | PLAY | SLIDE | 0.02 … 0.3 | 0.06 | exp |
+| `drive` | DRIVE | DRV | 0.25 … 4 | 1 | exp |
+| `mode` | DRIVE | MODE | 0 HOT / 1 WET | 0 | switch |
+| `level` | OUT | LVL | 0 … 1 | 0.68 | pow |
+
+Presets:
+
+- `classic-squelch` — Saw, RES high, a medium sweep: accents bark and stack
+- `round-line` — Low RES and little sweep: a round bassline under a mix
+- `square-bounce` — Square, short decay, moderate RES: a hollow rubbery line
+- `wet-scream` — WET mode driven, RES near the top: screaming accents
+
 ## bass
 
 **Fretwire**, instrument, 4 voices.
@@ -964,11 +990,11 @@ Presets:
 | `noise-level` | MIX | NOISE | 0 … 1 | 0 | pow |
 | `hpf-cutoff` | HPF | CUT | 20 … 15000 | 20 | exp |
 | `hpf-resonance` | HPF | PEAK | 0 … 2 | 0 | linear |
-| `cutoff` | LPF | CUT | 20 … 18000 | 700 | exp |
-| `resonance` | LPF | PEAK | 0 … 2 | 0.7 | linear |
+| `cutoff` | LPF | CUT | 20 … 18000 | 1200 | exp |
+| `resonance` | LPF | PEAK | 0 … 2 | 1 | linear |
 | `drive` | LPF | DRV | 0.25 … 4 | 1 | exp |
-| `lpf-mode` | LPF | MODE | 0 HOT / 1 WET | 0 | switch |
-| `env-amount` | LPF | ENV | 0 … 8 | 4.8 | linear |
+| `lpf-mode` | LPF | MODE | 0 HOT / 1 WET / 2 K35 / 3 OTA | 2 | switch |
+| `env-amount` | LPF | ENV | 0 … 8 | 2 | linear |
 | `level` | VCA | LVL | 0 … 1 | 0.44 | pow |
 | `age` | VCA | AGE | 0 … 1 | 0.4 | linear |
 | `amp-hold` | AMP ENV | HOLD | 0 … 20 | 0 | pow |
@@ -978,8 +1004,8 @@ Presets:
 | `amp-release` | AMP ENV | REL | 0.01 … 10 | 0.13 | exp |
 | `filter-delay` | FLT ENV | DLY | 0 … 10 | 0 | pow |
 | `filter-attack` | FLT ENV | ATK | 0.001 … 10 | 0.014 | exp |
-| `filter-decay` | FLT ENV | DEC | 0.01 … 10 | 0.14 | exp |
-| `filter-sustain` | FLT ENV | SUS | 0 … 1 | 0.18 | linear |
+| `filter-decay` | FLT ENV | DEC | 0.01 … 10 | 0.3 | exp |
+| `filter-sustain` | FLT ENV | SUS | 0 … 1 | 0.2 | linear |
 | `filter-release` | FLT ENV | REL | 0.01 … 10 | 0.11 | exp |
 | `mg-freq` | MG | FREQ | 0.1 … 20 | 3 | exp |
 | `mg-wave` | MG | WAVE | 0 … 1 | 0.5 | linear |
@@ -988,16 +1014,36 @@ Presets:
 | `mg-cutoff` | MOD | MG>CUT | 0 … 4 | 0 | pow |
 | `eg-pitch` | MOD | EG>PIT | 0 … 1 | 0 | pow |
 | `portamento` | MOD | PORTA | 0 … 10 | 0 | pow |
+| `ms20-c1-src` | CORD 1 | SRC | 0 OFF / 1 FENV / 2 AENV / 3 AENV- / 4 MG / 5 MG PUL / 6 S&H / 7 NOISE / 8 KBD / 9 VEL / 10 PRESS / 11 VCO1 / 12 VCO2 / 13 VCA OUT | 0 | switch |
+| `ms20-c1-dst` | CORD 1 | DEST | 0 OFF / 1 PITCH / 2 VCO2 / 3 PW / 4 HPF CUT / 5 LPF CUT / 6 VCA / 7 MG FREQ / 8 FLT IN | 0 | switch |
+| `ms20-c1-amt` | CORD 1 | AMT | -1 … 1 | 0 | linear |
+| `ms20-c2-src` | CORD 2 | SRC | 0 OFF / 1 FENV / 2 AENV / 3 AENV- / 4 MG / 5 MG PUL / 6 S&H / 7 NOISE / 8 KBD / 9 VEL / 10 PRESS / 11 VCO1 / 12 VCO2 / 13 VCA OUT | 0 | switch |
+| `ms20-c2-dst` | CORD 2 | DEST | 0 OFF / 1 PITCH / 2 VCO2 / 3 PW / 4 HPF CUT / 5 LPF CUT / 6 VCA / 7 MG FREQ / 8 FLT IN | 0 | switch |
+| `ms20-c2-amt` | CORD 2 | AMT | -1 … 1 | 0 | linear |
+| `ms20-c3-src` | CORD 3 | SRC | 0 OFF / 1 FENV / 2 AENV / 3 AENV- / 4 MG / 5 MG PUL / 6 S&H / 7 NOISE / 8 KBD / 9 VEL / 10 PRESS / 11 VCO1 / 12 VCO2 / 13 VCA OUT | 0 | switch |
+| `ms20-c3-dst` | CORD 3 | DEST | 0 OFF / 1 PITCH / 2 VCO2 / 3 PW / 4 HPF CUT / 5 LPF CUT / 6 VCA / 7 MG FREQ / 8 FLT IN | 0 | switch |
+| `ms20-c3-amt` | CORD 3 | AMT | -1 … 1 | 0 | linear |
+| `ms20-c4-src` | CORD 4 | SRC | 0 OFF / 1 FENV / 2 AENV / 3 AENV- / 4 MG / 5 MG PUL / 6 S&H / 7 NOISE / 8 KBD / 9 VEL / 10 PRESS / 11 VCO1 / 12 VCO2 / 13 VCA OUT | 0 | switch |
+| `ms20-c4-dst` | CORD 4 | DEST | 0 OFF / 1 PITCH / 2 VCO2 / 3 PW / 4 HPF CUT / 5 LPF CUT / 6 VCA / 7 MG FREQ / 8 FLT IN | 0 | switch |
+| `ms20-c4-amt` | CORD 4 | AMT | -1 … 1 | 0 | linear |
 
 Presets:
 
 - `acid-bass` — Squelchy resonant bass — play low, ride LPF CUT
 - `aggressive-fill-lead` — Short aggressive lead / fill
 - `buzz-lead` — Bright buzzy lead with a touch of MG wobble
+- `k35-round-bass` — K35 bass: full low end, a gentle peak, short EG2 sweep
+- `k35-scream` — K35 filters both peaked and driven: the early MS-20 scream
 - `mild-acid-bass` — Mild acid, not techno
 - `minimal-8th-bass` — Minimal 8th-note spoken-verse bass
 - `nasal-sneer-lead` — Nasal sneering mono lead
 - `octave-night-bass` — Faster octave night-drive bass
+- `ota-singing-lead` — OTA LPF, peak high and clean: the later MS-20, symmetric and singing
+- `patch-feedback-growl` — VCA OUT back into FLT IN: the OTA growls and sings as the note swells
+- `patch-fm-metal-bass` — VCO2 a fifth up into LPF CUT at audio rate: a metallic, clangy bass
+- `patch-sh-bubbles` — S&H on LPF CUT, clocked by the MG: random bubbling under a peaked K35
+- `patch-tracking-lead` — KBD on LPF CUT for full tracking, MG on PITCH for a slow vibrato, PRESS opens it
+- `patch-vowel-sweep` — FENV on HPF CUT with both filters peaked: a talking band-pass sweep
 - `portamento-highway-lead` — Smooth portamento highway lead
 - `resonant-funk-bass` — Resonant funky bass with filter bite
 - `rising-filter-fx` — Overdriven rising-filter FX sweep
@@ -1620,7 +1666,7 @@ Presets:
 
 - `parallel-squash` — Parallel compression: the dry signal beside a crushed, saturated copy. AMOUNT brings the copy up under it.
 - `harsh/bit-guitar` — Lows cut, crushed to 8 bits at 11 kHz, into a clean amp: aliased, zippy guitar.
-- `harsh/cross-bass` — Justice-style bass: a clean sub lane, and a dirty lane cut below 250 Hz, crushed to 12 bits, overdriven, through a Brit 4x12, slammed by a FET and taped.
+- `harsh/cross-bass` — Justice-style bass: a clean sub lane, and a dirty lane cut below 250 Hz, crushed to 12 bits, overdriven, through a bright Brit 4x12 and slammed by a FET.
 - `harsh/four-slot` — Distructor's layout: distortion, filter, chorus and amp in four slots, each slot's mix on a knob.
 - `harsh/gated-fuzz` — A starved fuzz into a fast gate: notes sputter and cut off dead.
 - `harsh/kodiak` — Big and gnarly: a clean low-passed sub lane under a lane of octave-down, dynamic folding, RAT and OTT. SIZE is the sub, GNARL the dirt.