Johann Cohen-Tanugi, let me know if you'd like to see any further validations on this ticket.  I have created DM-30657 to add the doPerBandRefMagLim reference culling I've been using above (the option will be defaulted to False, so to make use of it a specific and deliberate config override in any given obs package will be required).  I may also explore a more dynamic culling based on the icSrc detection limits (but perhaps not on the short term).  And, more importantly, Eli is working on the real fix to this vulnerability on DM-30490.

Lauren MacArthur Thanks a lot, I think this is fairly complete. Here is just a hist of the checkCcdAstrometry scatter for all the DC2 logs (run2.2i for the cognizants)

So indeed the value of 0.1" seems to nicely divide into nominal and pathological behaviours. The "nominal" part is a bit weird, with a high tail hiding a ~gaussian centered at ~20  mas, but all in all this validates the guesstimates of 0.1" as a good starting value.

That histogram is quite telling! My only concern about using it to set the config threshold to force failures would be if the scatter coming out of the checkCcdAstrometry script is very different from the one computed from within the astrometry task (i.e. the one printed to the logs). They certainly “should” be very close, but it’s worth checking.

LGTM

`checkCcdAstrometry` is using all the sources present in the reference catalog and just applies a magnitude cut while the astrometric scatter reported by the astrometry task is probably only using the selected reference stars (no galaxies). When I wrote `checkCcdAstrometry` I determined that it was important to keep as much reference sources as possible.

Regarding the failures of the astrometry task, on top of the case where you get a too large astrometric scatter, there are also cases were you get a too small scatter. It correspond to cases were the fit becomes under-constrained. As far as I know this should not happen or only happen marginally with DC2 data because the geometry of the focal plane is perfect and we do not need to fit high order polynomials.