fit
Fits each count once per generator and keeps, for each count, the candidate the ranking criterion prefers.
A generator only supplies the starting partition for a non-convex search, so the choice of generator can decide the result: on skewed samples the generators fail in different places (Jenks spends its cuts in the tail, Quantile is scale-free, HistogramValley finds gaps in density). Running several and keeping the best by criterion makes that choice a much smaller decision, at the cost of fitting each count once per generator. The results record which generator won each count.
A separate name rather than a fit overload, so that a plain fit() call stays unambiguous.
Parameters
the counts to fit
the generators to try at each count; by default Jenks, Quantile and HistogramValley
improves each generated partition
chooses one family per group