should min and max promote their arguments?

[StefanKarpinski]

On the one hand, if we don't promote, as we do now, min and max are prone to nasty union return types. On the other hand, if we promote, the returned value will not, in general, be one of the input values, but rather, a promoted value that is equal to it.

[JonasRauch]

I actually just had a case in my AD experiments where I wished min and max would promote the arguments. If you extend a number type T and want the new type S to interact cleanly with the old one, it is much nicer if promotion is done. Before I wrote explicit methods min/max(S,T) methods that did this, the vectorized functions, e.g. min(Array{S}, T) returned Array{Any} instead of Array{S}, which sucked.

So just to get all the vectorizing magic for Number types easily for new types, promotion would be good imho.

[StefanKarpinski]

This is a good argument. We've been on the fence about this for a long time. I really think that making min/max more well-behaved with respect to types is the right thing to do. That means promoting.

[pao]

I ran into the same thing with the existing ode23(), which I worked around by float()ing the result in 60c762c. It would be nice not to have to deal with that explicitly.