You should include even the rare cases in your model, since your model should approximate the data as well as practicable. Your GLTs that you actually care about will not test the beta coefficients from these rare cases.
A technical problem arises when you have 100% accuracy in a subject. Then the rare cases vanish, and the regressors for those cases become all zero. In that case, you can either omit these cases -- which has the unfortunate effect of making that subject's output file have a different layout than the other (less perfect) subjects' files, which makes going to the group analysis level a little harder -- or you can run will all zero regressors using the -allzero_OK flag. In the latter scenario, the beta weights for the all zero regressors will be set to zero. Since they wouldn't be used in your GLTs, the results should still be meaningful.
Hope this is clear.
P.S.: At this time, 3dREMLfit does NOT allow all zero regressors, unlike 3dDeconvolve.