ElevenLabs [laughs] tags break caption-sync QA gates

With ElevenLabs v3/v4 audio tags, a line whose TTS text starts with a non-verbal tag like [laughs] or [sighs] produces real audio (often 0.5-1 s) before the first aligned word. A QA check comparing silence-detected voice onset with the first word's timestamp then reports a large mismatch (e.g. -855 ms) even though word-level captions are correct. Exempt or downgrade lines whose tts begins with a bracketed tag, or measure onset from the alignment's first character instead of from silencedetect.