Sure, but to take an extreme example, suppose you have a division operator that fails to check for division-by-zero. Then your operator works in 100 * (1 - 2^63)% of all cases. That gives you 18 nines and sounds pretty convincing but it's hardly the point.
In the same way I'm wondering how meaningful it is to say that an algorithm works for 99.49% of all doubles, sampled uniformly. In other words, how could I use that benchmark to make a judgement about using grisu3?
In the same way I'm wondering how meaningful it is to say that an algorithm works for 99.49% of all doubles, sampled uniformly. In other words, how could I use that benchmark to make a judgement about using grisu3?