velxio/frontend/src/__tests__/buzzer-metronome.test.ts

297 lines
12 KiB
TypeScript

/**
* Buzzer — metronome quality
*
* A real use case for the buzzer: an Arduino sketch driving it as a metronome
* (analogWrite/Timer PWM, an accent on the down-beat). This exercises the audio
* SCHEDULING that makes a metronome usable — and guards it against regressions.
*
* It drives the buzzer's PWM handler with a metronome sequence (onset → note-off
* pairs carrying their simulated timestamps) against a CONTROLLABLE audio clock,
* and records the resulting Web-Audio schedule (one oscillator per click). Then
* it asserts the qualities a metronome needs:
* - even onset spacing (steady tempo)
* - one oscillator per click, no overlap, strictly monotonic
* - correct pitch per metric level (accent vs beat)
* - bursty per-frame delivery is absorbed (still even)
* - a tempo change re-locks immediately and cleanly
*/
import { describe, it, expect, beforeEach, afterEach, vi } from 'vitest';
import { PartSimulationRegistry } from '../simulation/parts/PartSimulationRegistry';
import '../simulation/parts/ComplexParts';
// ── Controllable audio clock + schedule recorder ─────────────────────────────
let clock = 0; // seconds; we advance it to emulate real time passing
type Ev = { kind: 'start' | 'stop'; when: number; freq?: number };
let sched: Ev[] = [];
function mockOscillator() {
const o: any = {
type: '',
frequency: { value: 440, setValueAtTime: vi.fn(), linearRampToValueAtTime: vi.fn() },
connect: vi.fn(),
disconnect: vi.fn(),
onended: null,
start: (w: number) => sched.push({ kind: 'start', when: w, freq: Math.round(o.frequency.value) }),
stop: (w: number) => sched.push({ kind: 'stop', when: w }),
};
return o;
}
function mockGain() {
return {
gain: { value: 0, setValueAtTime: vi.fn(), linearRampToValueAtTime: vi.fn() },
connect: vi.fn(),
disconnect: vi.fn(),
};
}
class MockAudioContext {
state = 'running';
destination = {};
resume = vi.fn();
close = vi.fn();
get currentTime() {
return clock;
}
createOscillator() {
return mockOscillator();
}
createGain() {
return mockGain();
}
}
beforeEach(() => {
clock = 0;
sched = [];
vi.stubGlobal('AudioContext', MockAudioContext as unknown as typeof AudioContext);
});
afterEach(() => {
vi.unstubAllGlobals();
});
// ── Metronome driver ─────────────────────────────────────────────────────────
const OCR2A = 0xb3;
const TCCR2B = 0xb1;
// OCR2A value for a target frequency (CTC, prescaler 64): f = 16e6/(2*64*(OCR+1)).
// Integer division, matching the firmware's `F_CPU/(2*64*freq) - 1`.
const ocrFor = (freq: number) => Math.floor(125000 / freq) - 1;
const ACCENT = ocrFor(1500); // 82 -> 1506 Hz
const BEAT = ocrFor(1100); // 112 -> 1106 Hz
function setupBuzzer() {
const sim: any = {
cpu: { data: new Uint8Array(512).fill(0), cycles: 0 },
pinManager: {
onPinChange: vi.fn().mockReturnValue(() => {}),
},
};
let pwm: ((pin: number, dc: number, timeMs?: number) => void) | null = null;
sim.pinManager.onPwmChange = vi.fn().mockImplementation((_pin: number, cb: any) => {
pwm = cb;
return () => {};
});
const cleanup = PartSimulationRegistry.get('buzzer')!.attachEvents!(
{ addEventListener: vi.fn(), removeEventListener: vi.fn(), playing: false } as any,
sim,
(name: string) => (name === '1' ? 11 : null),
);
return { sim, cleanup, hit: (ocr: number, simMs: number, clickMs = 25) => {
// onset: OCR set, PWM duty > 0 at simMs
sim.cpu.data[OCR2A] = ocr;
sim.cpu.data[TCCR2B] = 0x04; // CS22 -> prescaler 64
clock = simMs / 1000;
pwm!(11, ocr / 255, simMs);
// note-off one click later
clock = (simMs + clickMs) / 1000;
pwm!(11, 0, simMs + clickMs);
} };
}
// Pull the recorded notes (a start paired with its following stop).
function notes() {
const out: { on: number; off: number; freq: number }[] = [];
for (let i = 0; i + 1 < sched.length; i++) {
if (sched[i].kind === 'start' && sched[i + 1].kind === 'stop') {
out.push({ on: sched[i].when, off: sched[i + 1].when, freq: sched[i].freq! });
}
}
return out;
}
describe('Buzzer — metronome quality', () => {
it('plays an even metronome with an accent on the down-beat', () => {
const { hit } = setupBuzzer();
const beat = 500; // ms — quarter notes @ 120 BPM
for (let i = 0; i < 8; i++) hit(i % 4 === 0 ? ACCENT : BEAT, i * beat);
const ns = notes();
expect(ns.length).toBe(8);
// even spacing (steady tempo)
for (let i = 1; i < ns.length; i++) {
const gap = (ns[i].on - ns[i - 1].on) * 1000;
expect(Math.abs(gap - beat)).toBeLessThan(5);
}
// one oscillator per click, no overlap, monotonic
for (let i = 0; i < ns.length; i++) {
expect(ns[i].off).toBeGreaterThan(ns[i].on);
if (i > 0) {
expect(ns[i].on).toBeGreaterThan(ns[i - 1].on);
expect(ns[i - 1].off).toBeLessThanOrEqual(ns[i].on + 1e-6);
}
}
// correct pitch per metric level
expect(Math.abs(ns[0].freq - 1506)).toBeLessThanOrEqual(2); // accent
expect(Math.abs(ns[1].freq - 1106)).toBeLessThanOrEqual(2); // beat
});
it('absorbs bursty per-frame delivery (onsets even in sim time, jittery in wall time)', () => {
const { sim, cleanup } = setupBuzzer();
void cleanup;
let pwm!: (p: number, dc: number, t?: number) => void;
sim.pinManager.onPwmChange.mock.calls; // noop ref
// re-grab the callback captured during setup
pwm = sim.pinManager.onPwmChange.mock.calls[0][1];
const beat = 250; // even sim spacing
// wall-clock delivery jitters by ~one 60fps frame each onset (a real burst)
const wobble = [0, -15, 12, -10, 14, -13, 10, -8, 0, 11];
for (let i = 0; i < 10; i++) {
const simMs = i * beat;
sim.cpu.data[OCR2A] = i % 4 === 0 ? ACCENT : BEAT;
sim.cpu.data[TCCR2B] = 0x04;
clock = (simMs + wobble[i]) / 1000; // jittery wall time
pwm(11, 0.3, simMs);
clock = (simMs + 25 + wobble[i]) / 1000;
pwm(11, 0, simMs + 25);
}
const ns = notes();
expect(ns.length).toBeGreaterThanOrEqual(8);
// despite the wall-clock jitter, scheduled onsets stay even (de-jittered)
for (let i = 2; i < ns.length; i++) {
const gap = (ns[i].on - ns[i - 1].on) * 1000;
expect(Math.abs(gap - beat)).toBeLessThan(20);
expect(ns[i - 1].off).toBeLessThanOrEqual(ns[i].on + 1e-6); // no overlap
}
});
it('re-locks immediately and cleanly when the tempo changes', () => {
const { hit } = setupBuzzer();
let t = 0;
for (let i = 0; i < 5; i++) {
hit(BEAT, t);
t += 500;
} // 120 BPM
for (let i = 0; i < 6; i++) {
hit(BEAT, t);
t += 250;
} // jump to 240 BPM
const ns = notes();
expect(ns.length).toBe(11);
// no overlap / no backward note across the change
for (let i = 1; i < ns.length; i++) {
expect(ns[i].on).toBeGreaterThan(ns[i - 1].on);
expect(ns[i - 1].off).toBeLessThanOrEqual(ns[i].on + 1e-6);
}
// settled to the new rate within a beat of the change (no drift/creep)
const lastGap = (ns[ns.length - 1].on - ns[ns.length - 2].on) * 1000;
expect(Math.abs(lastGap - 250)).toBeLessThan(10);
});
// Regression guard for the maintainer's review (PR #220, comment 4671821612):
// a melody / continuous tone is consecutive tone(pin, freqN) calls with NO
// noTone() between pitches — back-to-back nonzero OCR writes, each firing the
// PWM handler with duty>0 and no intervening note-off. The old code overwrote
// activeOsc on every pitch change without stopping the previous node, so
// oscillators were "started but never stopped" — they stacked and played
// forever. The monophonic guard must REPLACE the live note instead.
it('replaces rather than stacks oscillators on a melody / continuous tone', () => {
const { sim } = setupBuzzer();
const pwm = sim.pinManager.onPwmChange.mock.calls[0][1] as (
p: number,
dc: number,
t?: number,
) => void;
// A little tune: pitch changes with no note-off between them (legato).
const melody = [523, 587, 659, 698, 784, 659]; // C5 D5 E5 F5 G5 E5
const step = 200; // ms per note
melody.forEach((freq, i) => {
sim.cpu.data[OCR2A] = ocrFor(freq);
sim.cpu.data[TCCR2B] = 0x04; // CS22 -> prescaler 64
clock = (i * step) / 1000;
pwm(11, 0.5, i * step); // square-wave duty>0, pitch change, NO dc=0
});
// End the tune with a noTone — releases the final note.
clock = (melody.length * step) / 1000;
pwm(11, 0, melody.length * step);
const starts = sched.filter((e) => e.kind === 'start');
const stops = sched.filter((e) => e.kind === 'stop');
// Every oscillator that started is stopped — no orphans left ringing. This
// is the exact failure the maintainer saw ("started but never stopped: 6").
expect(starts.length).toBe(melody.length);
expect(stops.length).toBe(starts.length);
// One start per note, at the right pitch, in order. Compare against the
// CTC-reconstructed pitch (what the firmware's integer OCR actually yields),
// not the nominal note — same round-trip the buzzer's getFrequency does.
const heard = (freq: number) => Math.round(125000 / (ocrFor(freq) + 1));
starts.forEach((s, i) => {
expect(s.freq).toBe(heard(melody[i]));
if (i > 0) expect(s.when).toBeGreaterThan(starts[i - 1].when); // monotonic, no backward
});
// Pitch is read FRESH per note (not stuck on the first onset): the tune rises
// then falls, so the heard sequence is non-constant and tracks the melody.
const heardSeq = starts.map((s) => s.freq);
expect(new Set(heardSeq).size).toBeGreaterThan(1);
expect(heardSeq).toEqual(melody.map(heard));
// Bounded overlap (replacement, not stacking): each note is released close to
// the NEXT note's onset — not smeared to the end of the tune. Pair each start
// with its own stop (interleaved start/stop/start/stop… once the guard fires).
const notesSeq: { on: number; off: number }[] = [];
for (let i = 0; i + 1 < sched.length; i++) {
if (sched[i].kind === 'start' && sched[i + 1].kind === 'stop') {
notesSeq.push({ on: sched[i].when, off: sched[i + 1].when });
}
}
expect(notesSeq.length).toBe(melody.length);
for (let i = 0; i < notesSeq.length - 1; i++) {
// old note ends as the next begins (≤ a release tail past the next onset)
expect(notesSeq[i].off).toBeLessThanOrEqual(notesSeq[i + 1].on + 0.01);
expect(notesSeq[i].off).toBeGreaterThan(notesSeq[i].on); // positive duration
}
});
// A melody that ends WITHOUT a noTone() — the real "sketch loops tone() and
// never calls noTone()" pattern. Correct Arduino semantics: a tone() plays
// until noTone() or the NEXT tone(), so the final note must keep ringing. The
// guard must release exactly the SUPERSEDED notes (one stop each) and leave the
// last one sounding — not orphan the middle notes, not cut the last one short.
it('releases superseded notes but leaves the final note ringing (no trailing noTone)', () => {
const { sim } = setupBuzzer();
const pwm = sim.pinManager.onPwmChange.mock.calls[0][1] as (
p: number,
dc: number,
t?: number,
) => void;
const melody = [440, 494, 523]; // A4 B4 C5
const step = 200;
melody.forEach((freq, i) => {
sim.cpu.data[OCR2A] = ocrFor(freq);
sim.cpu.data[TCCR2B] = 0x04;
clock = (i * step) / 1000;
pwm(11, 0.5, i * step); // pitch change, NO dc=0 — and no noTone at the end
});
const starts = sched.filter((e) => e.kind === 'start');
const stops = sched.filter((e) => e.kind === 'stop');
// Every note starts; only the superseded ones stop → exactly one note (the
// last) is still live. Pre-guard this was starts=3, stops=0 (all orphaned).
expect(starts.length).toBe(melody.length);
expect(stops.length).toBe(starts.length - 1);
});
});