LGTM

video/src/preflight.ts

@@ -1,32 +1,83 @@
 import { existsSync, mkdirSync, writeFileSync } from 'node:fs';
 import { join } from 'node:path';
 import { OUTPUT_DIR } from './config.js';
-import { storyboard } from './storyboard.js';
+import type { Storyboard } from './script.js';
+import { storyboards } from './storyboard.js';
 
 /**
- * Emit the narration script for the synth step.
+ * Emit per-storyboard narration scripts for the synth step.
  *
  * Synth (tts/synth.py) runs BEFORE recording, so it needs the full ordered
- * narration list — text + per-cue gaps — without depending on Playwright,
- * the dashboard, or auth. Walk the storyboard cues, write a flat manifest,
- * exit.
+ * narration list — text + per-cue gaps + voice config — without depending
+ * on Playwright, the dashboard, or auth. Walk each storyboard's cues, write
+ * a flat manifest under `output/<name>/narration-script.json`, then write
+ * an index manifest at `output/storyboards.json` so render.sh knows which
+ * storyboard slugs to loop over.
  *
- * The cue index in this manifest is the source of truth: the runner later
+ * The cue index in each manifest is the source of truth: the runner later
  * matches storyboard cues to measured durations by index.
  */
-function main(): void {
-  if (!existsSync(OUTPUT_DIR)) mkdirSync(OUTPUT_DIR, { recursive: true });
+// Em/en-dashes and ellipses make Qwen3-TTS produce dramatic pauses, sighs,
+// or audible breaths — the captions still render the original (unicode-rich)
+// text from the storyboard; only the synth input is sanitised.
+function normalizeForTts(text: string): string {
+  return text
+    .replace(/\s*[—–]\s*/g, ', ')
+    .replace(/…/g, '.')
+    .replace(/\.{3,}/g, '.')
+    .replace(/\s{2,}/g, ' ')
+    .trim();
+}
+
+function emitScript(storyboard: Storyboard): string {
+  const dir = join(OUTPUT_DIR, storyboard.name);
+  if (!existsSync(dir)) mkdirSync(dir, { recursive: true });
 
   const items = storyboard.cues.map((cue, cueIndex) => ({
     cueIndex,
-    text: cue.text.trim(),
+    text: normalizeForTts(cue.text),
     gapBeforeMs: cue.gapBeforeMs,
   }));
 
-  const manifest = { items };
-  const path = join(OUTPUT_DIR, 'narration-script.json');
+  // The voice block is consumed by tts/synth.py — see _resolve_reference and
+  // the cache check there for which fields invalidate cached audio.
+  const manifest = {
+    storyboard: storyboard.name,
+    voice: {
+      instruct: storyboard.voice.instruct,
+      language: storyboard.voice.language,
+      temperature: storyboard.voice.temperature ?? 0.6,
+      topP: storyboard.voice.topP ?? 0.9,
+      seed: storyboard.voice.seed ?? 42,
+    },
+    items,
+  };
+  const path = join(dir, 'narration-script.json');
   writeFileSync(path, JSON.stringify(manifest, null, 2));
-  console.log(`Wrote ${items.length} narration cues to ${path}`);
+  console.log(`[preflight] [${storyboard.name}] wrote ${items.length} cues → ${path}`);
+  return path;
+}
+
+function main(): void {
+  if (!existsSync(OUTPUT_DIR)) mkdirSync(OUTPUT_DIR, { recursive: true });
+
+  for (const sb of storyboards) emitScript(sb);
+
+  // Index for shell loops — each entry has every field render.sh needs to
+  // address per-storyboard outputs without re-parsing the TS source.
+  const index = {
+    storyboards: storyboards.map((sb) => ({
+      name: sb.name,
+      aspect: sb.video.aspect,
+      outputFps: sb.video.outputFps,
+      minDurationS: sb.video.minDurationS,
+      maxDurationS: sb.video.maxDurationS,
+      posterTimeS: sb.video.posterTimeS,
+    })),
+  };
+  const indexPath = join(OUTPUT_DIR, 'storyboards.json');
+  writeFileSync(indexPath, JSON.stringify(index, null, 2));
+  console.log(`[preflight] wrote storyboard index → ${indexPath}`);
 }
 
 main();