I don't know all the details behind Courgette, but it sounds very similar to Exediff. In my thesis I show that bsdiff (well, a slightly improved version of bsdiff which I never got around to releasing publicly) performed on par with or slightly better than Exediff; so I'm surprised that Courgette is getting this far ahead of bsdiff.
Based on the numbers they've published (a 10 line source code patch to a 10MB executable resulting in a 700kB bsdiff patch), it sounds like there's something weird going on which is breaking bsdiff -- the FreeBSD kernel is roughly the same size as Chrome, but normally for a small (e.g., 10 line) source code patch I would see a bsdiff patch of between 50kB and 100kB.
Beyond that, I really can't offer any more insight -- unless someone wants to hire me to spend a week looking at the old and new binaries and figuring out where bsdiff is going wrong. :-)