When I was doing performance examinations from localhost I found that CSS was mostly inconsequential if written at least vaguely efficiently and requested as early as possible from the HTML. By completely removing CSS I might be able to save up to 7ms of load time, but that was extremely hard to tell because that was well within the variance between test intervals.
Obviously trying to do an optimization designed to reduce the impact of latency between client <-> server is going to have no impact if you are testing on localhost where latency is already effectively zero.
That's not to say i think this optimization is neccesarily worth it, just that testing on localhost is not a good test of this.
https://github.com/prettydiff/wisdom/blob/master/performance...