I tracked this down to a single object in that tract that had the calib_psf_used flag set in the coadd catalogs, but for which the shape measurement failed. I wasn't selecting against shape flags, but nan's are excluded, so one array was ending up with N - 1 entries in a column that was being "broadcast" with one that had the full N entries. The reason we did not see this before was because the calib_* flags were not getting propagated to the coadd catalogs, so these plots were getting skipped.
While an object that was used in visit-level PSF modeling failing to get a shape measurement on the coadd is a bit surprising (and, indeed, rare...only a single case in three tracts), the reason this can happen is that the calib_psf_used flag only requires that 20% of the visits contributing have that flag set (so perhaps there were enough visits contributing that were funky in that area that the resulting object in the coadd was also too funky to get a proper fit).
I've fixed the problem by updating my flag culling in those plot functions.