I don't think it is possible to answer this question well without seeing the data.
I think it is extraordinarily unlikely that the stop word does not appear in the data, if it shows up in a topic. Perhaps you are somehow accidentally incorporating another corpus, besides your data? Another possibility is that a stop word (e.g. "run") does not appear in your data, but a related word (e.g. "running") does appear, and there is an algorithm that is doing trimming of words to their root words.
One thing you could try, to debug this weirdness, is to run your code on half your data, to see if these stop words still show up. If they do, run it on on the other half. Keep slicing up the data, and maybe you can narrow down to see exactly which part of your corpus is causing the "error".