Taking the ID from the source instead of generating a uuid() mid-pipeline is easy to miss in review, because it works until the first retry. Then every retry produces new IDs and dedup stops matching anything.