feat(videos): cycle highlight uses ElevenLabs alignment timestamps

[jjackson]

Summary

@jjackson: "we want the highlight to specifically transition when we say, learn, deliver, verify, pay onto each highlighted area when we say the word"

The narration-synced highlight from #451/#455 picked a reasonable proportional position (word_index / total_words) but that's an estimate — it transitions roughly when the word is said, not precisely. This PR wires up actual per-character TTS timestamps.

Pipeline

1. voiceover.ts::synthesize — switched to ElevenLabs' /v1/text-to-speech/{voice_id}/with-timestamps endpoint. Same audio quality + voice settings; the response is JSON with audio_base64 + alignment.{characters, character_start_times_seconds, character_end_times_seconds}. The alignment is persisted in the sidecar JSON next to the mp3.
2. scripts/render.ts — reads the cycle beat's sidecar, calls wordStartSeconds(alignment, "learn"|"deliver"|"verif"|"paid"|"pay"), and threads the 4 numbers through Remotion props as cycleStepStartSeconds.
3. Root.tsx + Intro.tsx — Cycle checks frame / fps against the 4 timestamps; switches activeIndex on each crossing. Falls back to the word-index proportional path when alignment isn't available (Studio preview, or cached audio from before this PR).

Cache compatibility

The sidecar JSON used to live without an alignment field. synthesize() now treats those as cache misses so the audio re-synthesizes once per program ($0.01-ish in ElevenLabs cost) to backfill alignment. Re-renders after that hit cache as before.

Verified

render.log excerpt:

Cycle step timings (seconds into cycle audio):
  { learn: 0.975, deliver: 1.509, verify: 2.229, pay: 3.599 }

Frame extracts from a fresh chc/run-001 render (cycle beat starts at video t≈4s):

Video t	Cycle-relative t	Highlighted	Expected (per spoken word)
5s	1s	Learn	Learn (0.975s ≤ t < 1.509s) ✓
6s	2s	Deliver	Deliver (1.509s ≤ t < 2.229s) ✓
7s	3s	Verify	Verify (2.229s ≤ t < 3.599s) ✓
8s	4s	Pay	Pay (3.599s ≤ t) ✓

Each transition coincides with the spoken word's onset.

🤖 Generated with Claude Code