feat(review): say when somebody else is reviewing a case

[haribo]

Summary

• ADR 0005, at last (say when somebody else is already reviewing a case #95): a case is held exclusively by its reviewer — opening claims it, a 30-second heartbeat keeps it, leaving lets go (SPA navigation and pagehide keepalive both), and a silent lock expires on its own after two minutes (OZALID_LOCK_WINDOW, lazy — expiry is read, never written)
• Occupancy is its own axis: the case state never moves on a claim; GET case exposes held {by, name, since}, the page wears a neutral eye pill, the reader's carousel goes inert with "X is reviewing this case — read-only until they let go"
• A held case takes no verdict but its holder's: SaveReview, comment and recording judgments answer 423 naming the holder; the dev's moves — track, deliver, discard — pass untouched (§7 already protects the bytes)
• Claiming is one atomic upsert: takes a free or expired lock, renews the caller's own, and somebody else's live hold answers who — safe under the parallel race
• Migration 00021 (case_locks), exercised Up→Down→Up; the expiry window answered in product.md §11 (validated)

Test plan

• Design tests named in §3.2: TestAHeldCaseRefusesAnotherReviewersVerdict, TestASilentLockExpiresOnItsOwn, TestTwoReviewersRaceForTheLock — the race truly parallel
• Carousel spec: inert pair, the holder named, no emit under a hold
• e2e named in the design: "a held case reads the same and refuses the verdict (say when somebody else is already reviewing a case #95)" — two real signed-in reviewers, 423 asserted on the API, release on leaving verified
• Full gates: gen-check, be-check, be-test (12), db-test, fe-check, fe-test (110), fe-test-e2e (64)
• Screenshots from the second reviewer's seat, both themes, validated

Existing test modified (flagged): the #70 expired-session e2e resumes the verdict as another account — it now releases the lock first, since a held case rightly answers 423 to anyone but the holder.

Closes #95