Even if you ignore the problem of self-modifying code, a completely static translation of x86 binaries is impossible. Things get hard real quick when indirect branches are involved, so you have to rely on a runtime solution when ahead of time solutions won’t work.

It's both. Install-time translation, plus runtime emulation for code that the system didn't get to translate (JIT, newly compiled applications, runtime loaded or obfuscated code, applications running from non-local storage, etc).

What is the difference supposed to be anyway?

It's user-mode emulation of x86_64 whether the translation happens just-in-time or ahead-of-time. AOT just can afford to spend more time optimizing out redundant emulation operations in each basic block.

It's not emulating privileged modes or devices, but neither is Microsoft's emulation (or user-mode QEMU for that matter.)

I think it does both, but I'm not familiar with what Geekbench 5 actually does and whether it can be fully translated ahead of time. Regardless the translation doesn’t seem to be able to make use of both sets of cores.