Point-in-Time Correctness for a Churn Model

You're building weekly churn-prediction training rows: for each customer, one row per week t, label = "cancelled within 30 days after t". The warehouse gives you these tables: customers (current snapshot, one row per customer, overwritten daily), support_tickets (event log with timestamps), and subscription_events (event log: upgrades, downgrades, cancellations, with timestamps).

1. Identify two specific ways naively joining customers would leak future information, and explain the mechanism.
2. Design a leakage-safe way to compute tickets_last_90d for the row at week t.
3. subscription_events contains the row event_type="cancelled". Is it ever safe to use this table to build features? What's the rule?