Which compensation strategy prevents duplicate writes in UML activity retries?

UML 2.5 activity diagrams model concurrent branches using fork and join nodes, and support compensation actions to undo partial work when errors occur. A precise design goal is to determine whether compensation handlers should be automatically retried following a write failure, or whether explicit user intervention is required to prevent duplicate original writes. The specification does not define a standard retry mechanism for write operations, leaving tool vendors to implement proprietary extensions with varying semantics. This ambiguity creates risk: automatic retry may re-execute the original write, while delayed intervention may leave the system in an inconsistent state without compensation executing.

Which approach preserves exactly-once semantics across diverse UML implementations? Does the specification need a normative retry policy for compensation actions, or should tool vendors define extension points? When should explicit intervention be mandated to avoid silent data loss?