Redirects could bypass the backend URL guard, and TTS playback waited for full buffering.
Recording could overlap or finish after teardown. Controls also ignored backend readiness.
Explicit formats and config-aware cache keys prevent stale audio after settings change.