gridPaddingDeviations

Further padding of the evaluation grid, in sample standard deviations.

The range alone is not enough. A kernel of bandwidth h spreads mass roughly 3h beyond the outermost observation, and the bandwidths of interest here run up to the one that merges the data into a single mode — which is of the order of the spread of the data, not of a small fraction of its range. Without this term the estimate loses visible mass off the ends of the grid at exactly the bandwidths the diagnostic cares about.

Mode counting is unaffected either way, since modes are interior. What this protects is density, which is offered for plotting and ought to integrate to one.