A multi-step repair can time out after changing one resource but before changing another. The team needs a recoverable partial-failure path. Which design is best?

1. Hide the timeout and always mark the resource healthy.
2. Assume any workflow timeout automatically reverses every service call.
3. Use bounded step deadlines, explicit failure handling, and appropriate compensating or resumable actions for committed side effects.
4. Restart every step with fresh business IDs and no check of previous changes.