Robert, it is because XOR (according to bpmn2 spec) does allow to be visited more than one time, what it does in converging mode is that it will not synchronize the tokens on it but move the flow as soon as the first token reaches it. It is possible and completely valid that at some point later another token will go through the xor converging gateway again and thus it does not cancel (skip) it.
Although I think that subprocess should be capable of canceling active nodes when it is terminating. Will check that soon and post back the result. If I would not do it within a week please ping me here ;)
Cheers