That explains why there will be noise and some degree of inaccuracy, and that the process is not trivial.

It doesn't at all explain why in a large sample there would be powerful unexplained systematic error about ChromeOS usage in particular.