--longitudinal flag only affects the anatomical processing that will end up generating the anatomical references and support mappings between MNI - T1w - BOLD runs.
Therefore, and similarly to the discussion we had recently (Multiple scan sessions some bad anatomicals), I think the
--longitudinal flag has little usefulness for fMRIPrep per se. You may want to use it if you are planning some longitudinal analysis of brain morphology and thus, you want FreeSurfer to run on that mode.
If that is not the case (longitudinal analysis of the anatomy), then there are two options. If the anatomy between sessions is not expected to change dramatically, then I would suggest running fMRIPrep without the
--longitudinal flag, letting it average the T1w images. If the anatomy changes a lot, then I’d say it is best treating sessions far apart as if they were different subjects.