feat(task-board): keep a card In Progress while its reviewer runs

[pedrofrxncx]

Summary

A card moved to In Review the moment its PR was opened, so it sat in the reviewers' lane for the whole time an agent was still working on it. In Review read as "waiting on a person" when nobody was waiting on anything.

Now:

when	lane
agent opens a PR	In Progress — review cycle opens
reviewer records a verdict	In Review
hand-off (no PR, stale preview, retries spent)	In Review, unassigned

Why this needed a column

The lane could not simply be changed. The review cycle — the boundary that decides which reviewer verdicts still count — was derived from the newest status_changed → in_review activity, which made the lane load-bearing: moving a card mid-review would reset the boundary and invalidate its own reviewer's verdict. That is the exact failure mode that stranded 13 prod cards holding an approval that could never merge.

So the cycle stops riding on the lane:

• review_cycle_started_at (migration 189) is the boundary, backfilled from the timeline so in-flight cycles survive the deploy.
• reviewCycleStart(activity, cycleStartedAt) reads the column, falling back to the old activity scan only for cards stamped before the migration. The option is a required field on every cycle reducer — a caller that forgot it would silently count a previous cycle's verdicts as current, so the compiler asks at each of the 11 call sites.
• inReviewPhase(item) replaces status === "in_review" in every automatic path, and listItemsPendingReview mirrors it (open cycle or the In Review lane — either alone leaves a real case unswept).
• openReviewCycleIfInProgress only stamps a card with no cycle open, so the PR-open hook, the MCP tool hook and the thread-finish backstop are all idempotent. That predicate is the same guarantee the old conditional flip carried.

Two regressions decoupling opens, closed explicitly

1. A failed reviewer run would have scheduled a Super Agent retry or parked the card on To Do, because the card is now In Progress. reactToFailedTaskRun skips a card with an open cycle — the reviewer has its own budget (spentAttemptsThisCycle), and the sweeper spends it.
2. closesOwnReview gated on the lane, so an author's own run could have marked its work Done while a reviewer was still reading it. It gates on the phase now.

Testing

Real Postgres (embedded), not an in-memory fake:

• migration up and down from an empty database
• the backfill against a seeded pre-migration card → picks the newest → in_review stamp, matching the old derivation
• bun test apps/api/src packages/shared apps/web/src → no new failures vs. the branch point (183 pre-existing, unchanged)
• bun run check, bun run lint, bun run fmt, knip clean

New/inverted tests: cycle idempotency + re-open-after-close + org scoping + the sweeper work list (real PG); inReviewPhase unit cases; and the tests that encoded the old lane behaviour are inverted rather than appended — advanceToDoneIfMerged, closesOwnReview, refuseIfMergePending, the PR-open reaction, and the implementer prompt.

Screenshots

None — no visual change. The card renders in the In Progress column, which is existing lane rendering.

---

Summary by cubic

Keeps a card In Progress while its reviewer runs: linking a PR now opens a review cycle instead of moving the card to In Review, so the board only shows In Review once it's a person's turn.

The cycle boundary is now review_cycle_started_at (migration 190) rather than the newest → in_review transition, backfilled from the timeline so in-flight cycles survive the deploy. inReviewPhase replaces status === "in_review" in every automatic path, and openReviewCycleIfInProgress is idempotent across the PR-open hook, MCP tool hook, and thread-finish backstop. A late PR link can no longer strand a card: a Super-Agent-owned card sitting In Review with no open cycle is pulled back to In Progress and stamped, while a card whose cycle is already open is left where it is.

Regressions closed by the decoupling

• A failed reviewer run no longer schedules a Super Agent retry or parks the card on To Do.
• closesOwnReview gates on the phase, so an author's run can't mark its work Done while a reviewer is still reading it.
• The thread-finish backstop treats the linked PR as what makes a card reviewable, so a card that attaches its repo at runtime stays In Progress instead of being parked In Review.
• The starting-sandbox status is only announced on a real pod boot, not on warm pod resumes.

^{Written for commit 13d3666. Summary will update on new commits.}