Sure, subtyping adds additional pain points (mostly separate from overloading, BTW). But I think this is mostly orthogonal to reified vs. erased generics. As it is, the .NET runtime supports reified generics and declaration-site variance (albeit limited to interfaces and delegates), and to my knowledge there's nothing stopping a different language/runtime from implementing reified generics even with usage-site variance (and Scala's type system only has declaration-site variance, anyway, AFAIK).
> The big ML-.NET type system mismatches have to do with things like pervasive overloading in .NET making type inference difficult, nothing to do with reified generics
You've hinted at a side-effect, but no, it has to do with subtyping [1], which naturally leads to co/contra-variance [2]. Or to put it bluntly, in a nominal type system that has sub-typing, Hindley-Milner type-inference is not only "difficult" but actually impossible. And because F# had to be interoperable with .NET, then it needed C# generics. Pity, because there's many things that F# misses, like Ocaml functors, structural typing, type-classes, higher-kinded types and the list can continue. And I just love how List.sum<^T> is defined [3] (static member, FTW).
[1] https://en.wikipedia.org/wiki/Subtyping
[2] https://en.wikipedia.org/wiki/Covariance_and_contravariance_...
[3] http://msdn.microsoft.com/en-us/library/ee353634.aspx