feat(shorts): cinematic v2 visual design + synced karaoke captions

- Fix critical bug: scenes were stacked without <Sequence> — every segment
  rendered simultaneously, last scene covering the whole video (root cause
  of 'too simple/ugly' visual)
- Scene variety: kinetic word-by-word title (hook), karaoke caption card,
  big-number stat scene (auto-detected: $12B, 65%, 3.2M), CTA center
- Karaoke captions synced to TTS: edge-tts SentenceBoundary timing via new
  scripts/tts-words.py (mp3 + sentences.json per segment); active sentence
  bright, others dimmed
- Rich background per scene: 4 color palettes (indigo/pink, cyan/teal,
  purple/pink, orange/red), drifting orbs, dot grid, ken-burns motion
- Real fonts: @fontsource Inter 700 + Montserrat 800 (display/captions)
- Glass card captions, progress bar per-scene, safe zone 900x1400
- Word-boundary truncation for stat labels (no mid-word cuts)
- Perf kept: no blur/SVG-noise (software renderer); ~2 min per 32s short

scripts/render-worker.mjs

@@ -17,7 +17,7 @@
  */
 import { PrismaClient } from "@prisma/client";
 import { spawnSync } from "child_process";
-import { existsSync, mkdirSync, writeFileSync } from "fs";
+import { existsSync, mkdirSync, writeFileSync, readFileSync } from "fs";
 import path from "path";
 
 const prisma = new PrismaClient();
@@ -39,6 +39,12 @@ mkdirSync(SCRATCH, { recursive: true });
 
 const REMOTION_ROOT = path.resolve(process.env.REMOTION_ROOT || "./src/remotion");
 const REMOTION_ENTRY = path.join(REMOTION_ROOT, "index.tsx");
+const TTS_PYTHON =
+  process.env.TTS_PYTHON ||
+  (existsSync("/home/code/.hermes/hermes-agent/venv/bin/python3")
+    ? "/home/code/.hermes/hermes-agent/venv/bin/python3"
+    : "python3");
+const TTS_WORDS_PY = path.resolve(process.env.TTS_WORDS_PY || "./scripts/tts-words.py");
 
 function sh(cmd, args, opts = {}) {
   const r = spawnSync(cmd, args, {
@@ -54,11 +60,18 @@ function sh(cmd, args, opts = {}) {
 
 function tts(text, outBase) {
   const out = `${outBase}.mp3`;
-  sh(EDGE_TTS_BIN, ["--voice", VOICE, "--text", text, "--write-media", out], {
-    timeout: 60_000,
-    env: { ...process.env },
-  });
-  return { file: out, durationSec: ffprobeDuration(out) };
+  const srt = `${outBase}.sentences.json`;
+  sh(
+    TTS_PYTHON,
+    [TTS_WORDS_PY, text, VOICE, out, srt],
+    { timeout: 60_000, env: { ...process.env } }
+  );
+  let sentences = [];
+  try {
+    sentences = JSON.parse(readFileSync(srt, "utf-8"));
+  } catch { /* fallback: single sentence */ }
+  if (!sentences.length) sentences = [{ w: text, start: 0, end: 0 }];
+  return { file: out, durationSec: ffprobeDuration(out), sentences };
 }
 
 function ffprobeDuration(file) {
@@ -81,6 +94,23 @@ function ffprobeSize(file) {
 
 function log(...a) { console.log(new Date().toISOString(), ...a); }
 
+/** Truncate at a word boundary (never cut mid-word). */
+function clipWords(s, max) {
+  if (s.length <= max) return s;
+  const cut = s.slice(0, max);
+  const sp = cut.lastIndexOf(" ");
+  return (sp > max * 0.5 ? cut.slice(0, sp) : cut).trim() + "…";
+}
+
+/** Detect a stat worth rendering as a big-number scene: $12B, 65%, 3.2M. */
+function extractStatData(text) {
+  const m = text.match(/(\$?\d+(?:\.\d+)?\s?(?:[BMK]%|billion|million|%))/i);
+  if (!m) return null;
+  const rest = text.replace(m[0], "").replace(/^[^A-Za-z0-9]+/, "").trim();
+  if (rest.length < 5) return null;
+  return { value: m[1].trim(), label: clipWords(rest, 70) };
+}
+
 /** Build segment text list from a YoutubeScript's structured fields. */
 function buildSegments(script) {
   const segs = [];
@@ -133,12 +163,12 @@ async function renderScript(script) {
   }
 
   // 3. Remotion timeline (frames)
-  const timeline = voiced.map((s, i) => ({
+  const timeline = voiced.map((s) => ({
     text: s.text,
     duration: Math.max(30, Math.round(s.durationSec * FPS)),
-    size: baseSegs.length <= 4 ? 80 : 68,
     label: s.label,
-    highlight: s.highlight || undefined,
+    sentences: s.sentences || undefined,
+    stat: extractStatData(s.text),
   }));
 
   // 4. Render silent video via Remotion CLI (props = segments JSON)

scripts/tts-words.py

@@ -0,0 +1,33 @@
+#!/usr/bin/env python3
+"""TTS with per-sentence timing (karaoke captions).
+
+Usage: tts-words.py <text> <voice> <out.mp3> <out.sentences.json>
+Writes mp3 + JSON [{w, start, end}] (seconds, relative to segment start).
+Uses edge-tts SentenceBoundary stream events.
+"""
+import asyncio
+import json
+import sys
+
+import edge_tts
+
+
+async def main(text: str, voice: str, mp3_path: str, sentences_path: str) -> None:
+    comm = edge_tts.Communicate(text, voice)
+    sentences = []
+    with open(mp3_path, "wb") as f:
+        async for chunk in comm.stream():
+            if chunk.get("type") == "audio":
+                f.write(chunk.get("data", b""))
+            elif chunk.get("type") == "SentenceBoundary":
+                start = chunk.get("offset", 0) / 1e7
+                dur = chunk.get("duration", 0) / 1e7
+                sentences.append(
+                    {"w": chunk.get("text", ""), "start": round(start, 3), "end": round(start + dur, 3)}
+                )
+    with open(sentences_path, "w") as f:
+        json.dump(sentences, f)
+
+
+if __name__ == "__main__":
+    asyncio.run(main(sys.argv[1], sys.argv[2], sys.argv[3], sys.argv[4]))