To me, this means that a speaker will slightly overshoot the mark as you come up to the top or bottom of a big impulse (e.g. bass drum hit) and will move beyond where it is being told to by the electricity flowing through its coil. To me this says that normalising at 98% will give you a little bit of headroom for the speaker to move.
I expect this has also occurred to the folks who design the speakers, and the rated maximum output of the speaker is such that transients don't bottom out the cone travel.
As for integer rounding problems, I'm no DSP theorist but surely the only normalisations free of rounding errors are ones where you multiply the whole lot by an integer, i.e. amplifying by 0%, 100%, 200%, 300%. So by that theory, normalisations near to an exact multiple (say 5%) should be less noisy than normalisations far from an exact multiple (say 25%). (People who know what they're talking about should feel free to destroy that argument using the power of maths if necessary.)
Peter