[thorwhalen]

Fixes #66.

store.update(annotation) replaces the annotation with that id atomically on every backend (added to the IntervalAnnotationStore protocol). Raises KeyError for an unknown id (add stays the way to create); the replacement may change interval or tier. Additive: nothing existing changes signature.

• SqliteStore: one BEGIN IMMEDIATE transaction updates the row and its R*Tree entry. On a connection shared between threads a reader would see the open transaction's state, so annotation reads (all, by_tier, __getitem__, _candidates, id lookup) now take the store's lock.
• MemoryStore: a same-interval replace is a single slot write; a move inserts the new copy before removing the old, so the id is never absent even to a lock-free reader. Writers and the snapshot readers share an RLock.
• MappingStore: overwrites the one mapping key (never absent in the backing mapping), then the in-memory copy, under the lock.
• PostgresStore: one UPDATE ... WHERE id AND owner/project statement (atomic in any isolation mode); TierOverlapError as in add.

Tests: tests/test_store_update.py (memory, mapping, sqlite): replace, move, unknown id, and the acceptance test, a reader thread against a replacing thread, same-interval and moving. Guard checked by mutation: swapping update for remove+add fails all six concurrency cases; removing the sqlite read lock fails both sqlite ones. Postgres tests are in tests/test_store_postgres.py (including a reader on a second connection). No Postgres server on the machine that built this, so those tests were not run here (they skip); the Postgres code follows the file's existing patterns.

Local: full suite with --doctest-modules: 902 passed, 5 skipped.

Adversarial review (fresh Opus subagent): no correctness defect in the update paths; Postgres checked by reading (placeholders, Range/Jsonb adaptation, KeyError via rowcount, ExclusionViolation). Two confirmed findings fixed in the second commit: (1) the reader thread could die and the test still pass, so it now records the reader's exceptions and asserts it ran; (2) the new RLock broke copy.deepcopy/pickle of MemoryStore/MappingStore, so __getstate__/__setstate__ rebuild the lock (test added). Also: MappingStore add/remove now take the lock, and a Postgres test for TierOverlapError from update. Left for follow-up (filed as an issue): MappingStore __setitem__/__delitem__ do not take the lock; iteration over the store while an update moves an annotation can still raise, as with add before.

🤖 Generated with Claude Code