Robert Lupton and I discussed this today, and came to the following conclusions:
- As several people have pointed out, we should expect there to be an offset between Kron and PSF and between Gaussian and PSF in the forced measurements. Both Kron and Gaussian have per-object apertures but do not account for the PSF, and the aperture corrections we're using do not work when that per-object aperture is calculated in a band (and hence with a PSF) other than the one we're photometering.
- What we don't understand is why we can see multiple loci (apparently corresponding to different choices for the reference band) in just the population of bright stars. Only faint objects near the i-band detection limit should have some other band as their reference band.
To debug, then, I think we want to identify some of these bright stars with non-i reference bands, look at images of them in all bands, inspect the reference-band-selection code, and generally figure out what's going wrong there.