The right one is from jira as it is what BPMN2 specification says and in 5.4 this is the behavior. However what you say about slower machines and error with gateway is very weird and can't really explain why it would matter. Can you provide test case for this?
Maciej, it seems there is a contradiction in what you say. The BPMN2 specification (according to the JIRA) says a terminating event node does not terminate the parent node. The jBPM docs v5.4 says the contrary. In my test with the last jBPM 5.4.0 engine, it works as the BPMN2 spec says. So I think someone has to modify the online doc? Thank you for you confirmation.
Antoine, yes, documentation is out of date could you please file a jira for documentation with reference to the docs section that needs to be updated?