Breaking:
  --openai-api-key and --elevenlabs-api-key are removed; set the
  OPENAI_API_KEY / ELEVENLABS_API_KEY env vars instead.

New:
  -o/--output-file writes the rendered audio to a file and exits;
  pair with --play to also play it. Output format is inferred from
  the file extension.
  Audio cache keyed by (backend, model, voice, text) skips repeated
  remote TTS calls; clear with --refresh-cache.

Renamed:
  --no-prompt is now --no-interactive (old name kept as a deprecated
  alias with a one-line warning).