Keep realtime tool responses ahead of hangup

[dimavrem22]

Executive Summary

Keeps accepted realtime consultations alive until their result-bearing response is fully delivered before local hangup.

Description

• Track consultations from acceptance through result submission.
• Correlate the next result response and require its final assistant transcript, response completion, and outbound audio drain.
• Defer local hangup with a bounded 30-second post-result deadline.
• Centralize socket and session ownership so local and remote termination close exactly once.

Reason

A local goodbye can race a background consultation and end the call before the caller hears the result.

Decisions

• Remote shutdown remains immediate.
• Timeout and error fallbacks remain response-bearing work.
• Failed responses release ownership safely.
• The bridge run lifecycle is the sole close owner.

Testing

• Full unit suite: 471 passed, 23 skipped.
• Contract suite: 5 passed.
• Ruff passed for changed files.
• Full-lifecycle regression covers delayed consultation, immediate two-step goodbye, final transcript, response completion, audio drain, timeout fallback, and exactly-once close.