Completely agree. Except on the implementation side. As a designer, I have been frustrated many times about useless (I mean, unused) features that I had to implement nevertheless, just for backward compatibility. Same as information the compiler has that is not given to hardware through the ISA. The best example is simply loading a pointer: to get that information in many RISC ISA, I'd have to either detect a pattern in the instruction stream or the memory accesses, while a dedicated instruction would have given me that information. I would have spent more silicon elsewhere, and saved development and verification time. That is a simple example, but I can assure you there are many. ISA does have an impact.
This does not mean I claim for a new way of delivering programs. You're right in saying there is more than the CPU. As enforcing recompilation is far more easily said than done. And again, I'm not asking for it. I'm just sharing my vision of a better delivery, seen from my designer perspective. It is just hypothetical, as is the thread. ;)