Two things landed in one commit that probably should have been separate. First -- SQLite in the coordinator for client auth and billing. The coordinator already knows who's querying what and how often. Adding metering there instead of bolting on a separate service was the obvious move. Embedded SQLite, no extra process, no network hop.
Second -- collections aren't created, they're discovered. When parquet files show up under a path, that path becomes a collection. No CREATE COLLECTION API, no admin panel, no provisioning step. Files appear, the system adapts. This wasn't a design decision so much as a consequence of how the source watcher already works. I just stopped fighting it.
The billing part was earlier than planned. The discovery part just fell out.