Does code cleanliness affect coding agents? A controlled minimal-pair study