afni.nimh.nih.gov

That is right. To a fair degree, afni_proc.py does not know (or attempt to figure out) the number of time points that have any stim at all in them (which themselves would overlap, since many studies have multiple stimulations within a single TR (cue, button press, etc).

In any case, the number of TRs per stim (class) refers to the number of non-zero time points in the regressor.
For descriptions, see:

gen_ss_review_scripts.py -help_fields

- rick