Once you develop a language, teach it to all your developers and establish a policy to punish deviations you can write very efficient programs - at near or exceeding asm performance (modern compilers are smarter than developers when it comes to CPU-specific optimizations).
Programming in raw C++ using all its features is probably not productive - the code will not be readable by anyone except Stroustroup himself.
Other languages such as Lisp also allow you to create your own language for specific problems, however they are focusing on expressive power and not on performance.
Also, with respect, it's fairly clear from your second paragraph that you don't understand the terminology being used at all. Which is okay, it's not your field. But, it makes your assessment of how complex this discussion is somewhat suspect.
I'd just be wary of using an extremely complex tool. From what I can tell, this "C++" looks like it could self-destruct at any minute, unless handled by an expert. Perhaps that makes the case even more strongly that you ought to filter your hires most strictly, but I wonder whether there might exist a simpler tool that could do the job with fewer risks.
"Yeah, this one time I double allocated a vtable on the stack of a doubly-linked binary hash."
And by the way, thanks so much for your work on OpenBSD (assuming you are the Ted Unangst of legend). I've been using it for years, and I appreciate the dedication of the folks who created it.
Perhaps C++ has some extraordinary benefits that I don't understand, which makes deleting the v-table before the method returns the stack to the heap worth the trouble [1]?
[1] I might have mixed up some of the terminology a little here. Feel free to correct me vigorously.