Yes, you are right, I didn't consider the checkcast instruction.
Yes, F# does not have this particular limitation, but has few of its own in different places. Most notably, it lacks HKT (arguably, because of CLR awareness of generics).
I suppose you meant boxing/unboxing. If so, yes, that's necessary for primitive types. As for regular objects, no casts are needed.
Regarding TCO, that's a well known limitation of the current Scala compiler. Hardly a "horror story".
Let me assure you, there are plenty of platforms where F# is nowhere to be seen. Thus, "can run on any platform [I care about]", which is a different set for different people anyway. For example, F# does not run on JVM, does it?