fix(task-board): a review cycle always ends in a verdict or a person

[pedrofrxncx]

Problem

Prod, org osklen, card OS-303 ("[GLOBAL][PERF] Habilitar cache de borda"): the reviewer thread reached completed and the card shows a red 0/1 with nothing able to move it.

The run never recorded a decision — its last parts are "Standing by for the background deno type-check…" and then finish. No review_approved / review_changes_requested on the timeline.

From there the card is a dead end:

• reviewerHandledThisCycle treats any non-failed terminal thread as a review ("a reviewer records its decision and completes"), so neither the sweeper nor the dialog poll re-dispatches.
• isSpentAttempt counts only failed/hung threads, so reviewerAttemptsExhausted is false and the hand-to-human path never fires.
• Only a verdict parks a card for a person (parkReviewedCardForHuman), so after feat(task-board): keep a card In Progress while its reviewer runs #6689 this card doesn't even sit in the reviewers' lane — it stays In Progress with an open review cycle, indefinitely, which is a quieter version of the same strand.

The board's 0/1 is correct throughout. There genuinely is no verdict. It just never changes.

Fix — three rungs, so the cycle always terminates

1. Ask. A reviewer thread that completed without a verdict is asked for one on its own thread (nudgeThreadTurn, the mechanism ensureReviewerCommented already uses) — its session still holds everything it reviewed, so this is cheaper and better-informed than a fresh reviewer. The prompt is narrow: record the verdict you already reached, request_changes if you could not finish, do not re-review, do not wait on background tasks.
2. Spend. If it still ends undecided, isSpentAttempt now counts it: verdictRecorded (from reviewCycleVerdicts on the current cycle) replaces "terminal and not failed" as the proof of a review. The existing MAX_REVIEWER_ATTEMPTS budget dispatches a second reviewer.
3. Hand over. Budget gone → the existing reviewerAttemptsExhausted → handTaskToHuman path, which parks the card In Review, unassigned.

The ask happens at most once per thread, fenced on the timeline (review_verdict_requested, migration 191) rather than in memory, so it survives a pod restart. An unanswered ask holds the attempt budget back for VERDICT_NUDGE_GRACE_MS (10 min), so the follow-up run's queue wait can't be mistaken for another silent completion.

The reviewer prompt also now says outright that ending without a decision is the one failure the run cannot leave behind.

Note on the extra read

enqueueEnabledReviewers now always reads the activity list. #6689 made the cycle boundary a column precisely to avoid that read — but the verdicts and the verdict-asks only exist on the timeline, so this path needs it regardless; the boundary still comes from reviewCycleStartedAt, with reviewCycleStart's activity fallback for pre-190 cards. One indexed query per swept card per 5 min, which is what this path cost before #6689.

Testing

• bun test apps/api/src/tools/task-board/enqueue-reviewer.test.ts — 42 pass, 8 new cases (undecided → asked once, failed/live never asked, timeline-scoped asks, grace window, asked-then-still-undecided → spent → hand-off).
• activity-actions.test.ts re-pointed at migration 191; notification-types.test.ts still holds — the new action is deliberately not a notification type, so NOTIFIED_ACTIONS keeps it out of the inbox.
• bun run check and bun run lint clean. The 10 remaining task-board test failures are pre-existing on main (they need Postgres).

Needs bun run migrate on deploy (CHECK constraint swap only).