Four teams, four versions, one production incident — and no way to answer "which model is actually live." Here's how enterprise ML teams use model registries to track lineage, gate approvals, and roll back with confidence when something breaks.