All of it!
No, seriously: I think a combination of your efforts would be best.
I see the charm of letting the 68k-JIT run on bare-metal, "below" the OS.
But how about allowing native ARM-code to bypass the JIT? (and execute it maybe on the second core?)
This way the result should be almost the same as running the JIT under ARM-AROS ... with just kind of a reversed default. Would this be possible?
And additionally a 68k-binary should be able to run in "high compatibility mode", that provides (some) chipset emulation. This should be selectable like via tool types or cli.
Too ambitious?