It's a matter of finding the right balance, and the size of the cache. The current core already requires software support for integer arithmetic and - quite like the 68060 - does not suffer from that because only exotic instructions require emlation that are rarely ever needed.
Consider an ATC of 256 entries, 4K per entry. This is already pretty small. The core could then address 1MB of memory without ever requiring a table walk. Put this into perspective to the typical Amiga program, probably at one tenth of this size.
Thus, the core would at worst initiate a table walk when loading a program.
Anyhow, this is all theory. I'm not saying that anyone should focus on this right now. There are more serious matters.
In an AMP context, your idea seems very promising. Especially as the architecture of the Amiga should facilitate this development.
And no, I have not forgotten the many discussions around a possible AMP based solution. In fact, I quietly waiting it.