If I understand correctly, there are two main concerns: speed and
portability. I think that they are both not that terrible.

How about this:
* For IEEE-754 hardware, we implement decimal/binary conversions, and
define the exact behaviour of floats.
* For non-IEEE-754 hardware, we keep the current method of relying on
the system libraries.

About speed, perhaps it's not such a big problem, since decimal/binary
conversions are usually related to I/O, and this is relatively slow
anyway. I think that usually a program does a relatively few
decimal/binary conversions.
About portability, I think (from a small research I just made) that
S90 supports IEEE-754. This leaves VAX and cray users, which will have
to live with a non-perfect floating-point behaviour.

If I am correct, it will let 99.9% of the users get a deterministic
floating-point behaviour, where eval(repr(f)) == f and
repr(1.1)=='1.1', with a speed penalty they won't notice.