In a clustered environment if a node fails everything should continue working as if nothing happened.
In a non clustered environment when jbpm fails, since the state is persisted it will continue from where it left off unless the failure occured while executing a work item i.e. while calling a web service, while executing a piece of code, as you mentioned between after the start of the work item and before the end of it. In those cases one will have to specify to the engine what to do with the persisted work items i.e. re-run them, complete them, do something based on some logic related to the work item etc.