The MaxDiff designer achieves excellent balance across a large number of versions, but in most cases it cannot produce perfect balance within each version (your case is a great example: asking 18 sets of 6 items each makes for 108 items shown in total. Dividing that by 35 shows that each item is shown 3.0857 times for a given person on average. So most respondents will see a given item 3 times but a few will see it 4 times). This happens necessarily, because your number of questions x number of items/question is not an integer multiple of the total number of items. And even if it were an integer multiple, we might be able to balance the "one way" item frequencies, but not the "two way" frequencies of how many times each item is shown with each other item (you only get this when a balanced incomplete block design exists for your exact problem and this is not always the case at all).
So there is necessarily imbalance between versions. Versions are assigned randomly across respondents when they START the survey and there is no provision for making sure we get the same number of completes per version, and this includes subgroups: you can, will and in fact did see different versions being shown more or less often within your subgroup of respondents.
This would be an important problem if you were analyzing you data with one of the simple count analyses described in the academic literature about MaxDiff: those count analyses work so well just because the experiments tend to include small numbers of items and the researchers have the luxury of making perfectly balanced designs - usually simple one block designs in fact. Unlike the counts and proportions utility estimates you may read about in the MaxDiff academic literature, when we use the multinomial logit statistical model to estimate utilities, it provides statistical control for the slight imbalances in the designs that respondents see.
Now, all that said, we know that there are version effects but research we've done finds that they are small, accounting for 2% to 5% of the variance in MaxDiff utilities, according to analysis done and published in our 2013 Conference Proceedings (see the paper by Chrzan and Hill here: http://www.sawtoothsoftware.com/support/technical-papers/100-support/proceedings/1426-proceedings2013