curious how schema evolution works with the lean types. if a column refinement gets tighter after you've already got rows in sql, do you surface that as a migration/check failure, or does it only blow up when decoding at runtime?