Storing a full candidate snapshot on every observation was the wrong call. The same node lists got copied thousands of times, about 22 GB on prod. Resolve endpoints live like path hops and drop the column.
