Component
Memoizes group fits by index range, so that a group range fitted once is never fitted again for the same sample.
This is the single largest saving available to the search. Refinement moves one cut at a time, so consecutive partitions share most of their groups, and a search over the number of components revisits ranges repeatedly. Fitting a group means running the whole estimator catalog over it, which dominates the cost of everything else in the pipeline.
Failures are cached alongside successes. A group range for which no family fits is expensive to discover and just as expensive to rediscover; caching only successes would leave the worst case uncached.
The cache is bound to one sample. It holds no reference to the data and does not verify that successive calls pass the same array, because an index range is only meaningful relative to a fixed sample; create a new cache for a new sample.
This class is not safe for use from multiple threads. Each worker in a parallel experiment constructs its own fitter and cache, which is also required because KSL's scoring models and metrics are mutable.
Parameters
the fitter to delegate to on a miss