That's already (one of) the main property that the algorithm is exploiting: That f64 is an horribly overkill representation for the data since it was generated

That's already (one of) the main property that the algorithm is exploiting: That f64 is an horribly overkill representation for the data since it was generated using a microsecond clock, leading to a lot of trailing zeros. So yeah, dropping more significant digits should lead to even better compression. But if we are going to be massaging the data based on what we know/need, there are better ways to go about compressing it. e.g. I'd expect that simple fixed-point quantization with a delta encoding would compress better than this. What I find cool about this is that it dynamically figures things out without having to deal with any priors.