Missing-data boundary
This case inserts actual nan fields into an otherwise valid growth CSV and
runs the native discover command. The command must reject the non-finite
observations. The result is intentionally not a fake imputation benchmark:
automatic missing-data policy is not currently exposed by this CLI path.