Idempotent DDL bootstrap for Ankusa.WAL.Postgres, run once at connection startup. Three tables:

• ankusa_wal — the log. seq is a BIGSERIAL; may have small gaps (a deduped or rolled-back row consumes a sequence value it never keeps) — harmless, since the contract only needs strictly-increasing seqs usable as a cursor, not perfect density.
• ankusa_wal_dedup — a permanent ledger, separate from ankusa_wal and never touched by truncate_through/2. This is what makes dedup survive WAL truncation: ankusa_wal rows get deleted once compacted to a segment, but the dedup key that row claimed is remembered forever, exactly mirroring WAL.DiskLog's persisted .dedup snapshot. Carries its own seq column (filled in by the same transaction that writes ankusa_wal) rather than looking it up via a join to ankusa_wal — that join would stop finding a key once its owning row gets truncated.
• ankusa_wal_cursors — one row per named reader cursor (:dispatch, :compactor), scoped by instance.

Every table is scoped by an instance column so one Postgres database can back multiple Ankusa.Instances — including the same instance name running on many independent BEAM nodes at once, which is the whole point: coordination happens through this shared, durable state, never through BEAM distribution.