So it is a McNemar generalized to more than a 2 X 2 framework, correct?

I believe your issue is that you have null values for counts. If you look at their examples, they never have null values. Also your log shows missing values, perhaps you end up with a zero in a denominators at some point in the calculation making a function undefined.

I went back and added a value in lieu of "0" in your dataset and the program runs fine. I am unsure how biased the results may be, but you could perhaps add a "0.5" to each count in the dataset to get the procedure to run. I am not theoretical enough to know how this may bias the results. It would definitely get you in the general ball park.

I ran their symptom example with and without the 0.05 correction and the results were similar. I think as long as the test-statistic is very large, you could have reasonable confidence the marginal distributions are different.