Nice measurments! The results are quite as expected, but without experiments you will never know, sometimes things are quite different than theory.
A thought that is not neccessarily correct: What is decreasing your luxvalues when using a collimating lens (apart from reflection/absorption losses) is that the beam intensity is not constant at all angles, but decreases with angle. So pre-collimating will project more of the lower-intensity edge of the beam on the lens, effectively decreasing the lens diameter and thus spot brightness. You showed that nicely in your last experiment with the pre-collimating lens a bit closer to the led.
About the Tondaj meter, I just re-read my luxmeters-post and realised: it over-reads cool white leds, but the 16% over-reading with a cool white XM-L (in a cheap dive-light) that I found in the test, later appeared to be a fake Latticebright led with a huge blue component in the spectrum (see Hannes’ post from yesterday), real XM-L’s have never been so blue so the over-reading of a genuine Cree will not get as high as 16%.