Checkpoint and reclamation

A checkpoint writes a new durable snapshot and WAL generation, then permits obsolete storage generations to be reclaimed. The database must be idle with respect to snapshot leases across all cooperating processes.

cargo run -p graphfusion --locked -- checkpoint --database ./demo-db

Database::checkpoint() returns Error::Busy when a statement or explicit transaction pins a snapshot. Schedule retries at idle points. Simply keeping already materialized Arrow result batches does not hold the snapshot lease.

DROP and graph replacement retire storage; they do not immediately delete files that a reader might still scan. Reclamation happens after the new manifest is durable and no active lease can reference the old generation. Orphan files from interrupted staging can be reclaimed as well.

Checkpoint is not Parquet compaction, an index build or a backup command. There is no automatic scheduler or supported live-directory copying backup protocol yet.