Dude, RTTI only works for classes with virtual functions or their derivatives.
Yes, I know. You're arguing that not all classes have vtables. I'm arguing that there's no reason that they can't, save for efficiency concerns. And if you're so concerned about efficiency that vtables are a burden, then you're already in a situation where an interpreted language is inappropriate for your needs.
Now you're arguing something different. Just because all classes can, doesn't mean that they do, or that a general purpose tool (like, say, a garbage collector) can assume that they do. It's not a question of efficiency, it's a question of what the language allows. If you're saying "everything has to have a vtable", then you aren't talking about C++.
Again, you're trying to set up a straw man: the question is not whether or not it's efficient to do type inference in C++ (for the record: it's NOT any more efficient in a dynamic language), the question is whether it's possible at all.
It is possible to do so, the original argument was incorrect, and you're trying to have an entirely different debate.