Keep realtime tool responses ahead of hangup

[dimavrem22]

Executive Summary

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

Description

• Track per-call asynchronous work from acceptance through result submission.
• Correlate owned responses and require final assistant transcript, response completion, and outbound audio drain.
• Defer local hangup with a bounded 30-second post-result deadline.
• Make remote shutdown immediate even while tool work is running.
• Make call and realtime socket closure idempotent.

Reason

A local goodbye can race background tool work and end the call before the caller hears the result. Remote termination must also remain prompt while work is outstanding.

Decisions

• Contact reads use a 30-second execution timeout and consultations keep their existing timeout.
• Remote shutdown never waits for pending tool promises.
• Late results after shutdown are ignored.
• Failed responses and timeout paths release ownership safely.

Testing

• Full suite: 833 passed.
• Typecheck, lint, and build passed.
• Regressions cover delayed contact read, immediate two-step goodbye, owned transcript and audio drain, immediate remote shutdown during consultation, late-result suppression, and exactly-once close.