Rust generics are early-typechecked. You can't call a method on a generic parameter unless you have bounds saying that you can. And, similarly, you can't substitute type parameters in places where the bounds don't agree. This means that there are much clearer API boundaries (no more documentation snippets saying "T must have a method .foo()" or "specialization FooTrait<T> must exist") and cleaner error messages, but less metaprogramming power. Rust is hoping to make up for that loss of metaprogramming power elsewhere, but in a vacuum C++ templates are more powerful (though they have other downsides)
(FWIW you can index BTreeMap or hashmap as well with []; the difference is that [] will panic if it can't find the item. entry() gives a reference to the entry which you can check for emptyness and update before pushing if necessary. In C++ most things get implicitly initialized so that's not a problem)
I've not seen much use of SFINAE, just enough SFINAE is a "core" type that makes it infect everything.
C++ templates are unique in that they let you do a lot more than that; for example, you can have a different implementation for one specific type. std::vector<bool> is implemented differently from any other std::vector<K>. You can also have templates over values, not just types; so you could define some T such that T<5> differs from T<6> in some way.
I don't know rust; does it allow stuff like this?
That said, I am also looking forward to true type-level integers; they're needed for the truly proper translation of fixed size arrays in the Fuchsia IDL (fidl).