Fix the video Voice choice; tighten the VideoSpec after review

[davior]

This PR has two parts, in separate commits: the Voice dropdown fix, then the fixes from the VideoSpec review.

1. The video dialog's Voice choice now works (731595f)

The Voice dropdown in the video dialog saved RenderOptions.voice, but nothing read it. Every render used the account's voice. The dropdown also only ever listed fal.ai voices. On an account with a Deepgram key, "auto" reads with Deepgram, so no choice in that list could have mattered.

• The dialog lists the voices of the engine that will read the video. A new helper, utils/renderVoices.ts, applies the same provider rule as the worker. "Account default" now names the voice it stands for, e.g. Account default (Haley — female, US (Flux)).
• The worker passes the choice on (video/worker.py:_render_voice) when it is a voice of the engine doing the reading. Otherwise the account's voice is used, so a choice saved under the other engine no longer turns into the model's first voice.
• The speech core honours a Flux voice id for Deepgram (routers/settings.py:synthesize_tts_bytes). Any other name still means the account's Deepgram voice. That includes read-aloud's fal.ai voice name, so read-aloud is unchanged.
• Re-rendering can't reuse old narration. The shot cache keys on the narration audio itself, so a new voice is never served cached shots from the old one.

2. VideoSpec review fixes (60c9d51, 0058942)

These fix what the adversarial review of the VideoSpec path confirmed. A separate skeptic reproduced every finding before I acted on it. Nothing that renders changes: the parity tests against the original walk still pass.

• Provenance:
  • When a quote or code block takes over a picture's section, the segment now records the picture's block, as does a dropped repeat of the same picture.
  • A heading that isn't read aloud keeps its block in its section.
  • The silent beat that holds a picture names that picture.
  • The spec's source records imported_at, and selection when only part of the note was imported.
  • Blocks are listed in document order.
• Schema:
  • The from / type tags are required, and compact dumps load back.
  • Dumps always use the spec's own spelling (from, in).
  • Transition, card-text and code-style settings refuse misspelt keys instead of silently using defaults.
• Compiler:
  • A clip whose file has gone missing is skipped, as its warning says, instead of becoming a shot that fails the render.
  • Now refused up front:
    • an intro that isn't first, or an outro that isn't last
    • an audio file used as a picture
    • a watermark or background that isn't an image
  • An orphaned anchor is named in the refusal.
• Doc:
  • §5.8 states the rule Notes renders by: a scene is cut into units at its overlaid beats.
  • The sounded-clip mapping, heading_attaches (marked as planned), stale line citations and the Background type are corrected.
  • New open decisions 12–14 for GVC.
• Confirmed but left alone, each with its reason given in the commit:
  • the pre-existing 500 on malformed blocks
  • the importer's internal labels
  • option URLs that aren't checked
  • overlay text kept as its own copy
  • unused title / shot_segments

Test plan

• Voice:
  • backend tests/test_video_voice.py (14 tests): the worker's choice per engine, Deepgram honouring a Flux id, and a render pinned to each engine
  • frontend renderVoices.test.ts (4 tests)
• VideoSpec: 15 new tests. Every one fails on main and passes here.
• pytest -q from backend/: 1,314 passed, 1 skipped (that skip was already there).
• npm test (208 passed) and npm run build from frontend/.
• CI green.

🤖 Generated with Claude Code

https://claude.ai/code/session_01PwaujUokXQAxnebBArpZuj