Your comparison of VC 2008 and 2010 is somewhat unfair. You compare a mature design for older platforms with a new technology built with the future in mind. VS 2010 codebase is probably crappier than the one of its predecessor. It happens all the time with major/complete rewrites. I am pretty sure that the bloat comes from this rather than the technology choice. I wonder if VS 11 is developed in .Net or not by the way.
Then, WPF (or any GUI for that matter) is not really a component you'll find in *really* time critical applications (think finance, web servers) and is probably not designed with the same objective in mind.
I guess that the guys in finance are interested as much as the guys in video games in the speed of their code.
The point I was trying to make is that the approach for increased performance is pretty much the same in native and manage worlds. It is hard to see where, in theory, the advantage of a purely native language is. Please don't get me wrong, I write C++ all day long, I am just trying to get a better picture of the problem.
It may be too late, but there's an important (IMHO) question that was not asked here. First some context.
If you look at what managed people are talking about in their conferences (notably YOW 2012), they are using the exact same subjects as native guys do. JVM and .Net guys are more than aware that, locality, caching and lock free structures are the future (and name your other thing talked about in the native world). Their talks are all about how they can achieve great (according to them) performance by only carring on the lower layers where it matters (that is interlocked operations, immutable data, array traversal etc... ) and few parts of the code actually do. In the case of the JVM (at least) they can have dynamic optimizations that are, for now, forbidden to native code (e.g. virtual function inlining).
The lesson seems to be : you don't need to use a native language for the performance critical part, you can hint the VM and the JIT on the difficult part and rely on them.
My questions is this point of view too optimistic for managed languages (and why because I don't think anybody reading this will say 'no' ?