You’re right, it’s pretty good! But maybe I should rephrase my question:
How WebPageTest can make such an assessment (ie, state that a gain of 40% is achievable as well known tools as jpegtrans, imagemagick can not even win one byte)?
It should be based on some specific tool/library that would be very interesting to know in order to try to achieve the above result (unless WebPageTest use JPEGMini? :)).
jpegtrans and imagemagick were trying to do lossless optimizations (stripping exif data, etc). WebPagetest checks jpegs for lossy savings by recompressing them at a quality level of 85 using libjpeg which is actually pretty conservative. Hand-tuning you should actually be able to get significantly better than that.