Honest question: Why would/could they not? There has been a ton of research published on extending HM type inference to support structural subtyping, row polymorphism, extensible records which very closely model many (but not all) of the concepts in Object Oriented systems.
That's actually a good question. I suppose that it would be possible in theory to try and reconcile the two, but I don't see that happening. And I don't know how much you can extend HM to support the full C++ featureset. I'd be surprised if it was possible.
I haven't seen the entirety of C++'s feature set added to an HM system, but even OCaml has an OO system that's fairly complete. Many new functional languages are adopting Daan Leijen's extensible records approach which can model OO fairly well. http://research.microsoft.com/pubs/65409/scopedlabels.pdf
All of these are extensions of ML that are shipped into production today, so it looks like more than just a theory at this point.
