Point-in-time feature correctness

Design the offline training-data generation layer of a feature store that must guarantee point-in-time correctness for a payment-fraud model. The training set is built from millions of labeled transactions, each with a decision timestamp; for every row you must join in feature values (e.g. 'merchant chargeback rate in trailing 30d', 'cardholder distinct-device count in trailing 7d') exactly as they would have been at that decision instant — never leaking a value computed after the label event. The same feature definitions also serve online at <10ms. Walk through how you store feature history, how the point-in-time join works at scale, and how you keep the offline and online code paths from drifting apart.