Thanks for the advice. I started to rerun my analysis using the --meshspace=caret framework, as you suggested. I was in the process of making everything in HCP space when I realized that for the problematic subjects, the alignment between diffusion/structural/freesurfer spaces was bad. I decided to add skull-stripping to the pipeline before the alignment, and now for all subjects the alignment is good, and the subsequent analyses also work.
Thanks for your help!