Yeah, that's effectively quantization, which will not work for general tolerance checks where you'd convert float similarity to int similarity.

There are cases where the quantization method is useful, hashing/binning floats being an example. Standard similarity checks don't work there because of lack of transitivity. But that's fundamentally a different operation than is-similar.