while most ICE's can't do that because they're too slow,

I hjave worked with Philips NZ (now a new name), Ceibo, Nohau, SILabs, Metalink and none of those have been anything but 100% real time.

pray tell which ICEs you refer to as being too slow.

e.g. where the CPU lives in programmable logic
in such cases I would say the solution is simple: use a core with a SILabs type JTAG interface.

Erik