The useful deliverable is the clean data plus the rejected data you can still inspect.
Forcing every bad row into the final output creates false confidence. The pipeline keeps exceptions visible and can provide roundtrip-verified split outputs for defined scopes.