Binding the approval to an action digest plus expiry is the bit most agent setups skip. Without it the approval basically covers whatever the model decides to do next. Do you recheck the budget at the same point as permissions, or earlier?
Both, for different reasons. I would check or reserve the budget before starting the operation so the harness does not begin work it cannot afford, then recheck it at the execution boundary alongside current permissions.
The second check matters when concurrent calls, retries, or earlier steps have consumed the remaining budget.
For an external effect, I think both gates should pass against the exact action being executed, otherwise the harness should return an explicit limit or permission outcome rather than continue.