YAY!!!! After a bunch of bug hunting and one conceptual change, I was able to verify that debinning can out-perform unbinned likelihood when the total number of signal events is indeed greater than the expected number of background events.

Even better than that, the performance is the best when using the covariant chisq
with an overall poisson penalization. Thus, I don't even have to try to justify
any crazy modifications to the chisq as a test statistic.

Now I can simply check that this is somewhat general, and see how well fitting
performs.

Changes to be committed:
modified: examples/discriminationMC.py