Sensei fixes for cleaner transcripts, better endings, and reliable completion notifications

Sensei, our AI conversation quality coach, found these problems by reviewing completed conversations and production signals.

• Live captions and new transcripts no longer show sound labels such as "[breath]" or "[cough]" as spoken words. A cough or breath on its own no longer becomes a conversation turn.
• If your agent closes a typed conversation by promising a scheduling link, it is now less likely to leave the participant waiting without that link.
• Completed conversations still send Slack and email notifications and run automations when a Trust Assessment cannot be generated. Automations that require a trust score are still skipped when no score is available.