Derive the report fields the model kept missing

[jserv]

A report was refused whenever the model returned an improvement plan in the wrong order, or phase tags that disagreed with that plan. Both of those follow from the plan itself, so the server derives them now and the model answers only for the mapping it alone knows, which feedback improvement each plan item addresses. What used to reach the candidate as "no evaluation" is an assessment they had earned.

The call around it was measured rather than guessed. Generation lands in six seconds at the median and past twelve at the tail, so the eight-second attempt timeout was cancelling healthy calls; thinking tokens come out of maxOutputTokens, so a thinking response arrives as truncated JSON; and the call budget was a product of two guesses, which let a report die with four of its six calls unspent after two consecutive 503s. Attempts get twenty seconds, the thinking budget is pinned to zero, and the pool is one number the deadline can pay for, spent by whichever loop needs it.

Separately, sanitizeReport bounded the report summary at the generic 300-character field limit while the agent validates it at 1200, and real summaries run 400 to 450 characters, so every candidate has been reading a committee summary that stops mid-sentence. Both halves are named constants now, held together by a test in the style this repository already uses for the integrity bound.

Verified with scripts/test.sh green and 74 live report generations against gemini-3.1-flash-lite across three transcript shapes: a near-silent candidate, an ordinary session, and a 92KB transcript. No failures. Two runs in that set hit consecutive transport failures and recovered on a later call, which is the case the old fixed retry allotment lost.

Not addressed here: the ten-minute wait in #31 is far outside these bounds and nothing in this branch explains it.

Refs #31

---

Summary by cubic

Reports that were otherwise valid no longer become “no evaluation” when Gemini returns improvement-plan items out of order or phase tags inconsistent with that plan. The server now normalizes those derived fields and preserves the assessment the candidate earned.

Bug Fixes

• Allows two semantic repairs and shares a five-call budget across repairs and transient retries.
• Raises report call and overall deadlines, disables Gemini thinking tokens, and delays the browser escape hatch to avoid abandoning healthy reports.
• Preserves summaries and fallback diagnostics up to 1200 characters instead of truncating them at 300.
• Includes validation failures in the final error instead of returning a generic schema message.
• Covers the report-generation timing and retry behavior related to 面試結束 Evaluation 產生時間過長 #31; the separate ten-minute wait remains unchanged.

^{Written for commit 51f4a12. Summary will update on new commits.}