It turns out Rosetta2 does not rely on special M1 hardware memory ordering at all:
https://twitter.com/never_released/status/153412764108259328...
https://twitter.com/marcan42/status/1534032369166225409
The baseline for translation is qemu-user, but other pure-software emulators like FEX-emu and Box86/Box64 can perform this translation significantly faster. So I think there's just ordinary performance engineering to close the gap.
The problem with x64 on arm64 is no longer if it's possible, but rather the performance penalty. As for whether I'll try to make a Rosetta-like compat layer using the hardware optimizations: not sure yet, I have bigger fish to fry before I get there.