another way to improve the results (you may already be doing this but I know some people don't) is making sure your two sample histograms are taken from exactly the same frames from each source.
I tried to implement the method from the paper Laserman mentioned by rotating the colourspace in avisynth and using repeated iterations of colourlike, but this resulted in further problems because the rotations had to be normalized to 255 which lead to further graininess when it was rotated back. (am I making any sense?)
In the end I decided the only way I would get suitable results was to write a new Avisynth pluggin based on the paper Laserman mentioned and using a method similar to colourlike but smoothing out the target histogram and including several random colour-space rotations within the pluggin. Basically you'll never get a really good result without considering multiple colour-spaces as there is a problem with crosstalk between the colour channels (which is something the author of colourlike was willing to admit)
Therefore I am currently learning C++ (I have previous programming experience but not in anything C based) hopefully I'll have something useful in a couple of weeks.
However, considering the limitations of colourlike your results are already quite good.