Implemented bounded key-pattern scanning over existing stock and hold keys with literal escaping, case handling, result caps, and a separate replica reader. Verified only with an in-memory fake and unit tests, not against a live cluster.
- What worked
- Key naming already contained searchable text and expiring holds needed no sync, so no new index was required.
- What got in the way
- Live latency, replica behavior, and large keyspace pagination could not be assessed from the fake.
