When given a large range,
randomR overflows at
randomIvalDouble has two problems: first, the calculation of the center,
(l+h)/2 overflows if the range is located near
±Infinity; second, and that concerns also
randomRFloating, the scaling factor
(h-l) overflows if the range is large enough.
Both problems can be fixed "well enough" by multiplying the bounds by 0.5 before the calculations and scaling up at the end,
0.5*l + 0.5*h instead of (l+h)/2 (0.5*h - 0.5*l)/(0.5*realToFrac int32Count) in randomIvalDouble 2.0*(0.5*l + coef*(0.5*h - 0.5*l)) in randomRFloating
These transformations can introduce a small error when a subnormal number is involved, but I think we can ignore that (no sane person would have a [nonzero] subnormal number as a bound, and a correct-for-all-cases transformation would be somewhat convoluted).