io.eezo.db.migrate

What the author chose to do with one drifted change.

Skip currently still records the new state in the snapshot, so a skipped change is not re-offered — see BACKLOG §28, which is the decision that has not been made.

The last step of a deploy: does the database this process is about to serve actually match the model it was compiled against?

The migration ledger and the migrations that fill it.

Every entry point takes a Tx, including the reading ones: applied ensures the ledger exists first, and creating a table is a write. Nothing here opens a scope of its own — a caller wraps the whole run in one transact, which is what makes a failed batch roll back including its ledger rows.