# `sampler`

<!-- begin:auto-docs -->
_(no submodules)_
<!-- end:auto-docs -->

## Purpose

`SamplerPlugin` is the universal polyphonic sampler. `loadSample(note, url)` fetches a raw `Float32Array` PCM buffer from a URL and registers it under a MIDI note key; `triggerAttack(note, velocity?)` finds the nearest loaded base sample and plays it back at the pitch ratio `2^((note − baseNote) / 12)` using linear interpolation. Concurrent voices are summed into the mono output channel, with deterministic voice stealing when the configured pool size is exceeded.

The plugin is a source (`io.inputs[0].count === 0`), so it slots into chains at position 0 — the chain's IO validator rejects placing it elsewhere. The processing path is allocation-free; only the voice pool array is mutated per block.

## Features

- Source plugin (`io.inputs[0].count === 0`), mono output (`io.outputs[0].count === 1`).
- One automatable parameter: `gain` (0..4, default 1) for master-out level.
- `loadSample(note, url, { signal? })` — fetches `url` as a raw Float32Array buffer; `signal` aborts pre-fetch (`AUDIO_FX.ABORTED`) or mid-fetch.
- `triggerAttack(note, velocity?)` — throws `AUDIO_FX.SAMPLE_NOT_LOADED` if no samples are registered; throws `AUDIO_FX.INVALID_NOTE` for `note ∉ [0, 127]` integer.
- `triggerRelease(note)` — switches the voice into release mode; one-pole exponential decay ends the voice cleanly.
- Voice pool (default 32, configurable via `initialMaxVoices` up to 128). Stealing is deterministic: the lowest-sequence voice is replaced.
- `getLoadedSamples()`, `getActiveVoiceCount()` accessors for diagnostics and meters.

## Usage: Basic Example

```ts
import { SamplerPlugin } from "@suaveplan/audio-fx-plugin-sampler";

const sampler = new SamplerPlugin();
await sampler.loadSample(60, "/sounds/piano-c4.raw");
chain.add(sampler);
sampler.triggerAttack(60); // play at native pitch
sampler.triggerAttack(72); // play one octave up
```

## Usage: Advanced Example

```ts
import { SamplerPlugin } from "@suaveplan/audio-fx-plugin-sampler";

const sampler = new SamplerPlugin({
    initialMaxVoices: 16,
    initialReleaseMs: 40,
});
await sampler.setup(
    { sampleRate: 48_000, channels: 1, blockSize: 512 },
    ctx
);

// Load a piano sample bank, key-mapped across the keyboard.
const ac = new AbortController();
try {
    await Promise.all([
        sampler.loadSample(48, "/sounds/piano-c3.raw", { signal: ac.signal }),
        sampler.loadSample(60, "/sounds/piano-c4.raw", { signal: ac.signal }),
        sampler.loadSample(72, "/sounds/piano-c5.raw", { signal: ac.signal }),
    ]);
} catch (error) {
    if (error instanceof AudioFxError && error.code === "AUDIO_FX.ABORTED") {
        console.warn("sample bank load aborted");
    } else {
        throw error;
    }
}

// triggerAttack(65) picks the nearest base (60), pitch-shifts +5 semitones.
sampler.triggerAttack(65, 0.8);
```

## API Reference

- `SamplerPlugin` (class extends `AbstractAudioFxPlugin`)
  - `new SamplerPlugin(options?: SamplerPluginOptions)` — throws `AUDIO_FX.INVALID_OPTIONS` on malformed options.
  - `loadSample(note, url, { signal? }): Promise<void>` — fetches and registers a base sample.
  - `triggerAttack(note, velocity?: number): void`
  - `triggerRelease(note: number): void`
  - `getLoadedSamples(): ReadonlyArray<SamplerLoadedSample>`
  - `getActiveVoiceCount(): number`
  - Inherited accessor for `gain`.

## Implementation Notes

The pitch ratio is `2^((note − baseNote) / 12)` — the standard equal-temperament pitch-shifting formula. We use linear interpolation between adjacent samples in the source buffer, which is the cheapest interpolator that avoids audible quantisation artefacts. For high-fidelity pitch shifts (e.g., spanning more than an octave from the base), callers should load multiple base samples spaced an octave apart so the nearest-neighbour lookup keeps the interpolation distance short.

Voice stealing uses a monotonic `sequence` counter on each voice. When the pool fills, we scan for the lowest sequence and replace it. The cost is O(maxVoices) per attack — for the default pool size of 32, this is 32 comparisons, well under the per-attack budget of a real-time engine. Faster algorithms (heap, ring buffer) would lower the constant factor but add bookkeeping complexity; the linear scan is the right MVP.

Release uses a one-pole exponential decay: `releaseGain *= 1 − releaseCoef` per sample with `releaseCoef = 1 − exp(-1/releaseSamples)`. When the gain falls below `1e-4`, the voice is removed from the pool. This is the same envelope-decay form used by the gate plugin and gives a perceptually-correct "soft release" rather than a click-prone hard cut.

`loadSample` mirrors the reverb's `loadImpulseResponse` pattern: pre-aborted signal throws `AUDIO_FX.ABORTED` before any fetch; signal aborted mid-fetch surfaces as the same error, leaving the existing sample map untouched. The signal field is built conditionally in `fetchInit` to satisfy `exactOptionalPropertyTypes` — `RequestInit.signal` is `AbortSignal | null`, and assigning `undefined` would be a type error.

The plugin output is mono (`io.outputs[0].count === 1`) so callers can chain a `MixerPlugin` or `StereoWidenerPlugin` downstream to place the sampler in stereo space. Spatial positioning is the consumer's responsibility — the sampler intentionally does not pan internally so a single instance can drive multiple voices that are panned differently downstream.