Data Validation Gates and a Feature Backfill

Your nightly training pipeline for a recommendation ranker has three stages: (a) build labels, (b) assemble features via the offline store, (c) write the training set. On Tuesday, an upstream team changed the device_type enum from {ios, android, web} to {iphone, ipad, android, web} without notice, and separately your team wants to change the definition of user_ctr_30d to exclude bot traffic.

1. Which validation checks would have caught the enum change, and what should the pipeline do when they fire?
2. Explain how to roll out the new user_ctr_30d definition without creating a training window that mixes both definitions.
3. What property must the feature-materialisation job have for the rollout in (2) to be safe to retry, and why?