End an interview when it is over

[jserv]

A thirty-minute mock interview was reported as taking over ten minutes to produce its evaluation, with no way to tell a slow report from a hung one, so the candidate closed the tab. The report call is not what took that long: it is bounded at 125 seconds and measures about six at the median, with thinking pinned off. The wait was the interview's own clock. Nothing could end a session except the countdown or the candidate's own button, so a conversation that genuinely finished at minute 18 of a 30-minute booking sat idle until the timer caught up, behind a spinner that said nothing about which of the two was happening.

The interviewer can now close a finished session itself, through the same control packet the browser and the server-side deadline already send, and refused until recorded evidence says the rounds actually ran. Alongside that: the ending overlay counts the wait instead of only spinning through it, a report that never lands offers the offline summary beside "leave the room" rather than making leaving the only exit, and a paused interview keeps its deadline instead of freezing until the server's two-minute grace expires.

The second half is the enhancement asked for on the issue: that evaluation happen during the conversation's natural pauses rather than all at the end. Each idle window long enough now spends a short call, off to one side of the room loop, on the stretch of transcript nobody has assessed yet; those notes reach the final report prompt beside the interviewer's own phase evidence. Both are added to the transcript rather than replacing it, because the communication score is read almost entirely from speech. To be plain about the trade: this buys what the report is written from, not when it lands, and session-wide token spend goes up rather than down even though the final call gets cheaper.

Verified with ./scripts/test.sh, the full offline gate: cargo fmt, clippy, the Rust suites, the Python and Node suites, ESLint, the generated-artifact drift checks, cargo-audit and the hook suite. All green. The frozen prompts under tests/golden changed on purpose and were regenerated with UPDATE_PROMPT_GOLDEN=1. Every load-bearing line was checked by reverting it and confirming a test noticed, which is how three fixes that had no coverage were found and given some.

Deliberately left out: the idle-window reviewer ships together with the prompt change that makes the interviewer's own phase evidence richer, so their separate contributions to report quality are not measurable from this branch. If that matters, the two are separable into their own commits.

Closes #31

---

Summary by cubic

The interviewer can now end a session when the conversation is finished, so candidates no longer wait out a timer after an interview is actually over. The ending overlay counts the wait, a report that never lands offers the offline summary, and paused interviews keep their deadline. Evaluation also gathers during natural pauses: short idle-window reviews send notes to the final report prompt, alongside rather than replacing the transcript.

Key changes

• Adds end_interview tool, gated on evidence that both coding rounds ran.
• Shows elapsed time on the ending screen and always offers the offline summary button.
• Paused interviews no longer lose their deadline.
• Idle windows over 8 seconds trigger a bounded review of unreviewed transcript, appended to evidence.
• Session-wide token spend increases, though the final call gets cheaper.

Testing

• Full test suite passes, including new unit and browser tests; golden prompts regenerated intentionally.

Closes #31

^{Written for commit 21bcdd2. Summary will update on new commits.}