If we are to use should_grp_score_cos(x,y) as a filter the the following
relationship must hold (from least to most expensive):
should_grp_score_len(x,y)
>= should_grp_score_cos(x,y)
>= grp_score(x)
should_grp_score_cos(x,y) wasn't holding up its part of the bargain, so
real data was used to generate a fudge curve to bring
should_grp_score_cos(x,y) results into the same space. Really this is a
terrible hack and the problem needs more thought. Evaluation of
should_grp_score_cos(x,y)'s performance benefit (given the relaxation of
the filter under the fudge curve) is sorely needed.