The trouble is that, if my theory is right that it's the same as a 5ms processing delay, the problem you will face when A/B testing is that (for example) getting the right shade of blue on a link is basically useless, but the difference between (say) this site with all the custom CSS and with minimal CSS (just enough to get the layout right) at all is significant - it's just that no one change tips it over the edge, it's the combination of a very large number of small things (distinguishing user name from comment, for example).