Hey Tom,
excellent results again! How do you read out the times? By “hand” or did you write some script?
I’m a bit surprised that the alignment after draft is not faster, since if you repeat any alignment, it will be much faster, even if you remove the time benefit by reusing feature detection. Maybe it would be more pronounced with larger scenes? Because with your handful of images one must consider that the results are probably not representative.
All this terminology can be looked up on various photogrammtry sites.
Reprojection Error: https://support.pix4d.com/hc/en-us/articles/202559369-Reprojection-Error
Mean/Median: http://www.diffen.com/difference/Mean_vs_Median
By setting the MAXIMUM reprojection error, you define a threshold above which RC rejects the points.
How high you allow the error to be depends on your needs for accuracy - higher error, more noise but also more possible features. So yes, an alignment can be better with a higher error, meaning there will probably be fewer and larger components. There are some threads about this - you can start with a higher error (you are correct about the scale, it is measured in pixels, btw.) and then tighten it down. That might give you more aligned images in the end.
Thanks for the praise! 
I’m still learning from many question, and your questions (especially the very specific, recent ones) contribute quite a lot…