Whenever you're ready, walk me through how you'd approach this.
How do you prevent data leakage when engineering features for a predictive churn model?
Explain how you would define the prediction timestamp, construct point-in-time-correct features, split the data, and validate the pipeline. Include production safeguards for future data availability, target-derived fields, customer-level overlap, and training-serving skew. Provide concise Python demonstrating leakage-safe preprocessing and evaluation without using information unavailable at prediction time.