Most of what jQuery does could be moved into standard browser APIs and then presumably implemented in native code. I feel that implementing it that way would certainly be significantly faster, because any performance loss from impossibility of trace-directed optimizations is outweighted by significant reduction of amount of crossings between native and JS code with completely different memory model. The fact that VM/native transition is expensive is exactly the reason why things like WebAssembly cannot directly manipulate DOM.

But this could not be implemented by writing C++ class that gets exported to JS code as $ when something that looks like jquery is loaded. In JS you can load jQuery and then monkey patch it, which will break when you use this kind of "transparent" optimization.