I have found some entropy in your example file around the 10 byte pivot, and believe that with a couple of days of work I can program up a model to search the entire solution space. That's the half-way milestone. From there, if there is 20% compression to be had, I will happily write up a C implementation of the compress/decompress pair. Either way you will know for certain that you have a solution, or there isn't one along these lines.
I will need access to several more sample files to improve the quality of the search space.