span-01 vs mercury-decide: same score, opposite failures

Two decision models scored the same F1 on my 23-case gate — but they missed different cases and only one stayed stable across days.