feat: dataset builder

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

tests/bench/dataset.test.ts

@@ -0,0 +1,37 @@
+import { test, expect } from "bun:test";
+import { mkdtemp, rm } from "node:fs/promises";
+import { tmpdir } from "node:os";
+import { join } from "node:path";
+import { buildSong, loadSongDir } from "../../src/bench/dataset.ts";
+
+test("buildSong writes all artefacts; loadSongDir reads them; stems sum to mix; idempotent", async () => {
+  const root = await mkdtemp(join(tmpdir(), "stsheets-ds-"));
+  try {
+    const dir = await buildSong("t1a", 1, root);
+    const files = ["song.json", "truth.gp", "mix.wav", "mix.mp3"];
+    for (const f of files) expect(await Bun.file(join(dir, f)).exists()).toBe(true);
+    const { song, truth, mix, stems } = await loadSongDir(dir);
+    for (const t of song.tracks) expect(await Bun.file(join(dir, "stems", `${t.name}.wav`)).exists()).toBe(true);
+    expect(truth.tracks[0]!.notes.length).toBe(song.tracks[0]!.notes.length);
+    expect(mix.channels.length).toBe(2);
+
+    // Oracle stems are the exact mix contributions: they sum to the mix (PCM16 rounding only).
+    expect(stems.size).toBe(song.tracks.length);
+    let maxDiff = 0;
+    for (let c = 0; c < mix.channels.length; c++) {
+      const m = mix.channels[c]!;
+      for (let i = 0; i < m.length; i++) {
+        let s = 0;
+        for (const a of stems.values()) s += a.channels[c]![i] ?? 0;
+        maxDiff = Math.max(maxDiff, Math.abs(s - m[i]!));
+      }
+    }
+    expect(maxDiff).toBeLessThan(2e-3);
+
+    const before = (await Bun.file(join(dir, "mix.mp3")).stat()).mtimeMs;
+    expect(await buildSong("t1a", 1, root)).toBe(dir);
+    expect((await Bun.file(join(dir, "mix.mp3")).stat()).mtimeMs).toBe(before);
+  } finally {
+    await rm(root, { recursive: true, force: true });
+  }
+}, 120_000);