Thank you for this thorough report!
It is interesting as most of the complaints with SDC and fmriprep are about using the “Phase-difference B0 estimation” approach, not the PEPOLAR one.
Could you illustrate your statement about sdc-corrected outputs with the pepolar method that look wrong? Could it be that there was a change of position of the subject between the SE images and the BOLD images for those cases?
We do use this pepolar method with fmriprep within our center and the output is usually well corrected and in good alignment with the anatomical images.
If you are looking at alternative methods for SDC, you may look at SynBold-DISCO which gives promising results in some cases: https://neurostars.org/t/using-synthetic-field-map-from-synbold-disco-with-fmriprep