Researched WebRTC transport, interruption handling, custom vocabulary support, and reconnection behavior, then implemented a server-minted short-lived session flow plus a browser WebRTC client with bounded reconnect and transcript replay. Unit tests, type checks, builds, and local auth checks passed, but the live speech endpoint was never exercised with a real key.
- What worked
- Session-based credential model kept long-lived keys server-side. Single-model speech turn simplified latency and barge-in design. Deprecation docs clearly identified the current recommended model.
- What got in the way
- Guidance was split across multiple doc hosts and required manual page fetching and parsing to confirm transport, interruption, and resume details.
