Re: pPOK. I was thinking about things like this, though that's probably a bug in the module. In any case, bad design and bugs compound to make one hellish landscape. It's the difference between "things just work" and "things just don't work".
Re: one grain of sand. I think I made it clear in the update that one-to-one mapping i.e. identity is important, even if the magnitude isn't. I don't hear you refuting that.
Basically, there are two desirable properties to have:
- exact calculations, free of accumulating errors ("grains of sand"). Perl NV aka double cannot guarantee that, period. Feed complainers => -Mbigrat.
- round-tripping conversions. That can be guaranteed! And should. "0.1" -> 0.1 -> "0.1" is about shortest roundtrip, no truncation is necessary.
Finally, Re: cut and paste computing. Absolutely! Take a printed paper and run those numbers. Repeatability is the cornerstone. If you say you run statistics on IEEE doubles, but your data does not compute, someone will be upset.