Have you read this thread (Experimental support for Relion's Bayesian polishing in csparc2star.py - #67 by olibclarke)?
I wonder if it could be due to the orientation of the relion-extracted particles not matching the defocus angle - so effectively the astigmatism is wrong?
In reference to this:
- Reassign particles to motioncor2 mics, extract, refines to 3.0 Å with same settings
Was this using the particle CTF estimates (presumably derived from Patch CTF on Patch Motion mics) or did you rerun Patch CTF on the mc2 mics? If not, that might be worth doing (and then make sure to extract using the micrograph CTF estimates) - if the issue goes away after running Patch CTF on the mc2 mics, that would be informative.