[thorwhalen]

Addresses the three follow-ups of #68 (review of #67):

• MappingStore.__setitem__ / __delitem__ now run under the store lock, like add / remove / update, so an update racing one of them cannot bring a deleted key back into the backing mapping.
• MemoryStore.__iter__ (and so MappingStore) iterates a snapshot taken under the lock, and __len__ takes the lock, so a moving update can no longer raise "dictionary changed size during iteration". (MemoryStore is not a Mapping subclass and defines no items() / values(), so __iter__ was the only iteration path.)
• The Postgres TestUpdate tests were run against a real PostgreSQL server (all of test_store_postgres.py passes), and a moving-update-under-concurrency case was added (test_moving_update_under_concurrent_readers: the id is looked up by all() and by an interval query while it alternates between two intervals).
• Tests in tests/test_store_update.py: iteration under moving updates for both stores, and a delitem-vs-update consistency check on the backing mapping. All three fail on main and pass here.

Closes #68

🤖 Generated with Claude Code