Seedbox v0.1 + firmware, generic kit/synth, ROOT/SCALE pages

First commit of the repo. Contents beyond the v0.1 simulator and layout:

- Full Pico 2 W firmware (firmware/app, drivers, engine port, host test).
- BANK page and 404 mode A/B removed; learned pads play exactly what was learned.
- 404/MiniFreak renamed to KIT/SYNTH throughout (jacks, code, sim, docs,
  layout labels; wiring rebuilt).
- ROLL returns to page select after a re-roll; a freshly learned track turns on.
- LCD contrast 66.
- ROOT and SCALE pages: 21 scales, AUTO = the seed picks; global and sticky.

Intended music change: the scale table grew from 8 to 21 and AUTO picks from
all of them, so every seed's scale changed. Vectors regenerated with
npm run vectors, plus new key-override cases. Flash record is now version 2.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

docs/decisions.md

@@ -0,0 +1,47 @@
+# Design decisions
+
+Short record of why things are the way they are, so they don't get undone by accident.
+
+## Music
+
+- **Tracks are only on or off.**
+  - There's no auto mode. Bringing parts in and out is the performance (Lyon, "Techno Counterpoint").
+  - Every track is rendered every bar, so muting doesn't change the pattern.
+- **No break lane.** Chopping breaks is the M8's job. Seedbox makes rhythm and melody. AMENIZER is a generator that plays a break pattern on the kit, not a sample player.
+- **Energy moves hats only a quarter as much as the rest.** Before that, energy mostly meant busier hats.
+- **Rhythmic frames per seed.** Before frames, most seeds put the kick on 1 and the snare on 3.
+- **Sections are transforms of A**, not new patterns. The phrase hangs together.
+- **Drift is continuous overlapping layers**, not steps. Wide knob spans used to do little. Drift adds displacement and weirdness at the top end.
+- **AMENIZER keeps its own feel.** Its humanising and chops are tuned so that DRIFT past halfway isn't too extreme. The generic layers come in later and at half strength.
+- **ROOT and SCALE are global and sticky (2026-10-11).** They survive SEED, ROLL and slot loads, so a seed can be fitted to an existing idea. AUTO lets the seed pick from all 21 scales, which changed every seed's sound (vectors regenerated). Slots don't store the key.
+- **Integer-only engine** with a seeded RNG. This makes it deterministic and portable, and testable by golden vectors.
+
+## MIDI
+
+- **Kit pads are learned, not configured.** The bank/pad/mode tables were dropped. BANK was dropped too (2026-10-11): it only made sense for the SP-404MKII's bank layout, and the kit side should work with any sampler or drum machine. Re-learn to change kit.
+- **STOP sends CC 120 on the kit channels** because some samplers (the SP-404MKII) ignore All Notes Off. SPP isn't sent.
+- **The synth is in poly mode on one channel.** Bass, melody and pad share it.
+- **MIDI OUT follows the 3.3 V spec (CA-033)**: 10 Ω on TX, 33 Ω on 3.3 V. An earlier draft had them swapped; the independent review caught it.
+- **Generic kit and synth (2026-10-11).** Jacks, screens and docs say KIT and SYNTH, not 404 and MiniFreak. Those two are the test rig, nothing more.
+
+## Hardware
+
+- **6N137 for MIDI IN**, because that's the part on hand. It's pin-compatible with the 6N138 here except pin 7, which is the enable input and must be high: a link ties it to pin 8, and the 4.7 kΩ pin 7 resistor is gone. C1 (supply bypass) is required and sits right beside the chip. Never fit a 6N138 with that link in place: on a 6N138, pin 7 is the transistor base.
+- **3.5 mm TRS MIDI jacks (Type A)** instead of 5-pin DIN: the DIN jacks didn't fit the box. Type A TRS gear takes a plain stereo cable; DIN gear gets a Type A adapter.
+- **The Pico's USB end is at the stripboard edge** (its first pin row is strip A), so the box is powered and flashed through a slot in the back wall, with no extension cable.
+- **11 momentary track buttons in a diode matrix**, with the state shown on the LCD map. Latching or lit buttons weren't available locally, and LEDs would need more pins.
+- **No FILL button.** Dropped after trying the panel; fills still happen on their own in bar 4. The RUN toggle took the free corner of the grid (row 1 / column 4, outside the matrix), and GHOST moved into FILL's matrix slot (row 2 / column 4).
+- **Nokia 5110** instead of a 0.91" OLED, which was too small to read.
+- **Landscape panel** in an enclosure of about 80 × 150 mm.
+- **One stripboard** in a plastic box, Pico socketed, everything else panel mounted. No standoffs, screws or USB extension.
+- **GP0/GP4 links run under the Pico between its headers.** That keeps them from crossing header pins; it's the reason for the col-13 cut length.
+- **Pad names P1–P35**, not W…, to avoid clashing with strip W.
+
+## Process
+
+- **The stripboard was reviewed by blind agents** with no access to the design reasoning, as well as by `check-layout.js`.
+  - First review (30-row DIN board): 6 issues, all fixed, then clean.
+  - Second review (USB at the edge, TRS jacks): no electrical faults, 12 doc and mechanical findings, all fixed. The biggest was a link note that named the ground strip instead of strip Y. The rest: USB slot height, SEED above R4, and the panel tool defaulting to portrait.
+  - R4 now lies flat along strip X across a cut at X23, so a normal ¼ W part fits under SEED.
+  - The re-check came back clean.
+- **The firmware port happens in Claude Code** (this repo), against the golden vectors.