Unfortunately it's now well known that preprocessing does affect quite a lot the results, so that there is probably no easy or analytical answer here. I think your best bet is to do a differential testing approach here: try to use SPM preprocessing with various tweaks to reproduce fmriprep weird result, and try to use fmriprep with tweaked parameters avoid these weird result. It seems you already have some leads, did you try to enable distortion correction in SPM? To speedup your tests, you can preprocess and then denoise in CONN only the specific subjects with the weird results.

Hope this helps,
Best regards,
Stephen