Filed issue DM-2775 to deal with the fitting iteration.

Dominique Boutigny raised these additional points on HipChat in the CFHT reprocessing room:
1- I think that it is better to leave out the saturated sources
2- It would be better to be able to pass the list of bad flags as a configurable parameter. It would then be easier to test various combinations
3- I don't understand how you deal with rotations in matchOptimisticB.py

Here are my responses to Dominique's comments:
1 - You are right. I accidentally removed the step of culling non-good sources before running the WCS fitter. I will fix that. A test showed that by culling saturated sources we reached 18.6 mas in 3 iterations of the outer loop, which is fewer iterations than before. Culling sources that have been interpolated in their center, as Paul Price suggested, does just as well and is probably superior so I'll do that. That catches saturated sources, cosmic rays, bad pixels, etc.
2 - I would love to make this more configurable, but it's tricky because some tests rely on more than fields that are not flags. Instead I made SourceInfo configurable by subclassing. And of course for quick tests one can simply edit the code.
3 - This is in reference to the loop in MatchOptimisticBTask, and I am not happy about it either. Vello Tabur said it was a poor idea to match repeatedly with broader and broader parameters, but rather to tighten parameters after getting an initial match with loose parameters. I think the outer loop in AstrometryTask does a nice job of tightening parameters so I plan to try eliminating the loop in MatchOptimisticBTask.

I tried simplifying the loop in MatchOptimisticBTask but the results failed on CFHT visit=850587 ccd=21 (which is usually the one that fails if any of the CFHT demo CCDs fail). My conclusion: this fixing (3) will take much more thought and care. Best to leave it alone for now.

I tried further orthogonalizing the commits but gave up in frustration. I had already spent 4 hours cleaning up the commit history and I think it'll have to do.

Otherwise I think I addressed the concerns. The results are quite good on the CFHT data (better than our old task) but still need to be proven on more distorted fields before I am comfortable making the new astrometry task the default in pipe_tasks.