Rebuild a resumed session that keeps failing

[jserv]

Gemini Live can close a socket with 1011 and then close every socket resumed from its checkpoint 10 to 25 seconds later, so the interview spent all eight restarts resuming the same failing conversation and ended. A resumed socket that dies before it is healthy and debt-free now makes the next replacement rebuild from local state, and resuming stays off until a reply beyond the recovery briefing, or a socket that lives past a minute owing nothing, proves recovery. The reply to a recovery briefing no longer resets the restart budget, since every replacement asks for one, including when that reply goes through a tool call.

During the same loop the candidate's speech was not transcribed, so the idle timers, stamped only from transcripts, counted a talking candidate as silent and nudged them. Until the open socket has transcribed the candidate, their microphone level above room noise now counts as speech for those timers; once it has, transcripts alone count again, so room noise cannot hold nudges back.

A cold rebuild replays the owed prompt and the editor, so if the 1011 is triggered by that input rather than by the resumed conversation, rebuilding will not help. The log now separates the two cases: Gemini cold rebuild also failed before recovery appears when a cold socket fails too, and each replacement: line carries owed_chars= and editor_chars=. No prompt changes, so the interview contract bundle stays at 26.

Verified with ./scripts/test.sh (green; browser lanes, cargo-audit and shellcheck skipped locally for missing tools) and cargo mutants --in-diff over the Rust changes.

Close #198

---

Summary by cubic

Fixes resumed Gemini sessions that keep failing after a 1011 close, and stops nudging candidates who are talking but not being transcribed.

• A resumed socket that dies before it is healthy now makes the next replacement rebuild from local state; resuming stays off until a reply beyond the recovery briefing or a socket that stays up past a minute proves recovery. Health is judged by how long the socket stayed up, not the debt-adjusted age the restart budget uses, so a long-lived socket replaced while owing a reply is not blamed as a failed resumption.
• The reply to a recovery briefing no longer resets the restart budget, whether the briefing was sent at once or held for an unpause or the end of a thinking hold; a briefing answered with silence or cut off by the candidate ends the briefing's turn too, so the next real reply counts as recovery proof.
• Until a socket has transcribed the candidate, microphone level above room noise counts as speech for the idle timers; once transcribed, transcripts alone count again. A replacement socket must transcribe the candidate again.
• The log separates a resumed session that fails before recovery from a cold rebuild that also fails, and each replacement: line now carries owed_chars= and editor_chars=; a checkpoint is named only when resumption is actually used.

Migration

• No prompt changes; the interview contract bundle stays at 26.

Closes #198.

^{Written for commit 7742a47. Summary will update on new commits.}