Building coadds and running ap_pipe with many different settings revealed some inconsistent behavior. In summary, the PSF FWHM used to compare the template and science image appears to still be based in pixel units in some places, which is a bug when they don't have the same platescale. There are also bugs in the option to convolve the science image rather than the template, which you would want to do if the seeing of the science image was better than that of the template. One is that the kernel attributes appear to still be calculated for the case where the template is convolved. The other appears to be an incorrect noise threshold, but that will require further investigation.
This ticket covers the work spent investigating these issues, and the initial bug fixes.