fix: dataset layout data/<tier>/<seed>, mix tests, cli hardening

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

tests/bench/dataset.test.ts

@@ -35,3 +35,23 @@ test("buildSong writes all artefacts; loadSongDir reads them; stems sum to mix;
     await rm(root, { recursive: true, force: true });
   }
 }, 120_000);
+
+test("multi-track t1b: stems on disk sum to mix.wav", async () => {
+  const root = await mkdtemp(join(tmpdir(), "stsheets-ds-"));
+  try {
+    const dir = await buildSong("t1b", 1, root);
+    expect(dir).toBe(join(root, "t1b", "1"));
+    const { song, mix, stems } = await loadSongDir(dir);
+    expect(song.tracks.length).toBeGreaterThan(1);
+    expect(stems.size).toBe(song.tracks.length);
+    let maxDiff = 0;
+    for (let c = 0; c < mix.channels.length; c++) for (let i = 0; i < mix.channels[c]!.length; i++) {
+      let s = 0;
+      for (const a of stems.values()) s += a.channels[c]![i] ?? 0;
+      maxDiff = Math.max(maxDiff, Math.abs(s - mix.channels[c]![i]!));
+    }
+    expect(maxDiff).toBeLessThan(2e-3);
+  } finally {
+    await rm(root, { recursive: true, force: true });
+  }
+}, 120_000);