Fix Version/s: None
Component/s: ci_hsc, meas_modelfit, obs_subaru
When per-pixel variances are turned off, CModel likelihoods are computed without using the variance at all. This would not matter in a pure likelihood fit, but it means the prior and likelihood are not given the appropriate relative weights - and the relative weighting is not even consistent; it depends on the noise level of the image.
Since the typical effect of this is to make the prior much more informative, there is some danger that fixing this bug will cause other problems due to poorly-constrained fits. To avoid this, I'll add a configuration option to tune the relative weighting of the prior via a constant (which we could set to the typical variance level of the images to get behavior like what we have now without the inconsistency).
So, it turns out this change breaks ci_hsc: there's a check there that greater than 95% of the sources we marked as calib_psfUsed must have base_ClassificationExtendedness_value == 0 on the coadd, and with this change that number is only 94.3% (3 out of 530 sources were classified differently). I think I'm going to call that unnecessary strictness in ci_hsc rather than change the extendedness threshold, but we should consider revisiting the extendedness threshold in the RC processing.
The modified ci_hsc is now running successfully through jenkins, so I'm ready to merge. I'll leave it open a bit longer in case Paul Price wants to comment on the ci_hsc changes.
Love the plot!
Some trivial comments on the GitHub PRs.