Another great observation from Yusra AlSayyad! I’ll have to dig a bit deeper to confirm, but a cursory glance leads me to believe this “distance” is just taken from the persisted match catalog from the processCcd.py stage...i.e. your intuition was bang on. It may be best just to get rid of these plots altogether (clearly, I’ve never paid them much heed...)

Would you mind giving this a look when you get a chance.  Example plots for the 3 different calibrations:

Couple of questions on github. Looks good.

For another ticket:
The main metric that people will be looking at is the mean distance.To make it more meaningful, a median would be useful, and so would and so a SNR < 100 median.

I went ahead and added the median value to all the "distance" plots.  It has always been at least indicated on all the scatter plots as the dotted grey line (it's the median that is used for the clipping for the stats), but it is now explicitly printed on the "distance" plots (as this quantity is not symmetric about zero and is not expected to follow anything approaching a Gaussian distribution).   I left it off the other plots for the sake of minimizing clutter, but it can be added to any of them in a future ticket if they are deemed useful and worth the added clutter. Here are example plots with the median now printed to them:

A full suite of the COSMOS coadd and a single COSMOS visit can be perused at: https://lsst-web.ncsa.illinois.edu/~lauren/lauren/DM-18745/plots/

The sub-sample selection for the stats will be based on S/N as of DM-19189.  Just to be sure, did you mean the sign you indicated above (i.e. that you want a "S/N less than X" median in addition to a "S/N greater than X" one)?

I'll wait for another sign-off from you in case you want to comment on the added commit.

Confirmed via Slack that the metric of interest (median for S/N > 100) is on its way under DM-19189.