A good migration plan separates schema change, application change, backfill, and rollback. R1 can reason through that split in the planning phase, then prove it did not blur those boundaries while executing.
The verification layer matters here because migrations often look correct in code review while still being dangerous in runtime order. R1 can block at the gate when the rollback story is missing or the verification probes are too thin.
The point is not “agents can migrate databases.” The point is “the evidence needed for a careful migration is part of the run, not something humans reconstruct later.”