Find a happy system for reweighting

[SamStudio8]

Currently, reweighting works pretty well, apart from cases like this:

We end up "spending" the observed GA's on the first few rounds of path finding. Leaving us only able to select the unweighted GG option in future paths. This is technically correct, but we need a method that perhaps uses more delicate reweighting?

[SamStudio8]

etc.

[SamStudio8]

In a quick and dirty fashion, we can cap the ratio used to reweight the Hansel matrix and this yields more accurate results (on the order of 0.2-2%). This somewhat confirms the hypothesis that aggressive reweighting is likely to be the cause of the close-but-no-cigar results on some of our datasets.