A pipeline step occasionally fails with a transient service error. Repeating it is safe because its outputs are idempotent. Which approach best fits?

1. Use bounded retries with backoff and observable failure handling.
2. Validate expected data freshness and volume in addition to job exit status.
3. Use a date-scoped idempotent backfill strategy.
4. Declare explicit dependency and success conditions in the orchestration workflow.