I tracked down this issue. It turns out there were some inconsistencies in doing the transformations to the Principal Colors for the fit and wired coefficients (we do this for both PSF mags and CModel mags and they were getting intermixed). This is fixed here and we now also use the same points in all mean and stddev calculations (i.e. those that were used in the fit and survived the sigma cuts).
This is the pre-fix plot:
And here is the same post-fix:
Updated plot for all tracts can be found (for now!) in: https://lsst-web.ncsa.illinois.edu/~lauren/lauren/DM-18532/plots/color/
(Oh, and I should add that the issue revealed itself when I changed the flux-to-plot-against column from PSF to CModel. The default set in coaddAnalysis.py is also adopted in colorAnalysis.py...I may not have noticed otherwise!)