Multi-node PostgreSQL persistence for the MCP Tasks extension.

This adapter accepts an application-owned Ecto.Repo; it never starts a repository or runs migrations. Task snapshots and their serializable work descriptors are stored atomically as versioned JSONB. Applied events are committed to a separate ledger in the same transaction as the aggregate update.

Claims use PostgreSQL row locks, database time, opaque UUID tokens, exact expiry matching, and monotonically increasing generations. Recovery uses FOR UPDATE SKIP LOCKED, allowing independently supervised runners on several nodes to consume the same queue without a coordinator process.

Recovery remains at least once. Applications must deduplicate external side effects with Snodo.Extensions.Tasks.Work.idempotency_key.

create/4 refuses a task with {:error, {:capacity_exceeded, limit}} when the table already holds :max_tasks tasks (default 10,000) or the caller's scope already has :max_active_tasks_per_scope working or input-required tasks (default 100). Either may be :infinity. The counts are taken inside the creating transaction under a transaction-scoped advisory lock, so concurrent creations on several nodes cannot pass a limit together; setting both limits to :infinity skips the lock and the counts. An expired task counts until it is reaped.

reap/1 deletes at most :reap_batch_size expired tasks per call (default 500). A runner calling it every :reap_interval_ms therefore removes at most that many tasks per interval.