> I think the first approach gives a lot more room for the host to create optimizations, to the point we could see hardware with custom instructions to make WASM faster

Heh, there were literally CPUs with some support for the JVM! But it turns out that “translating” between different forms is not that expensive (and can be done ahead of time and cached), given that CPUs already use a higher level abstraction of x86/arm to “communicate with us”, while they do something else in the form of microcode. So it didn’t really pay off, and I would wager it wouldn’t pay off with WASM either.