I had assumed they were reversed because the phase encoding direction given in the json file for the fmap is opposite (i.e. i vs i-),
For phase-difference fieldmaps, that does seem likely to produce a sign-inverted estimate of the field, which would then lead to incorrectly expanding/contracting when applying correction.
We’re using fMRIPrep 20.2.7 as advised here after we were having issues with SDC in v22
If you wouldn’t mind trying out 23.0.2 for the fieldmaps that are working well, I would be very interested in seeing how much of the old behavior we’ve managed to recover.