Dear all (in particular nebula828):
We believe(!) we could identify the problem and have fixed it. The problem seems to be the side branch in the loop, which contains a non-executable/failing CSV writer. We don't think it is related to meta nodes etc (there is some race condition involved; it depends on which node gets executed first: the side branch node or the loop end). One very simple workflow that replicates the problem is attached (which is now also part of the testing suite).
Can you check whether that workflow is similar to yours and shows the same behavior you are encountering? If in doubt we can also make a nightly build available to you to test whether that fixes your problem.
Thanks,
Bernd