How html5-boilerplate helped drastically improving our website performance

The transfer time improved by 0.6s (first view) and 0.2s (subsequent view), but the total load time worsened by 9 seconds (first view) and 3 seconds (subsequent view).

What am I missing?

You do not measure "website performance" without doing many repeated (better yet, interleaved) trials! These statistics are meaningless without standard deviations (at the very least). I could perform two one-off trials and conclude that unplugging my keyboard improves the "performance" of your website by X% because there happened to be a fluctuation in network conditions.

You can do the same thing by using Jammit with Rails (http://documentcloud.github.com/jammit/) or mod_pagespeed with Apache (http://code.google.com/speed/page-speed/docs/module.html).

They both have the advantage that it's done on-the-fly and you don't need to do a build before deploying.