Here I've got some nice improvement from reordering the alphabet before BWT.
208107 on book1 with incomplete optimization.
http://ctxmodel.net/files/BWT_reorder_v1.rar
First, I tried to use some heuristic approach to compute the alphabet
permutation, but that failed miserably. (http://ctxmodel.net/files/BWT_reorder_v0.rar)
And then I remembered that specific alphabet not only affects the BWT,
but bitwise CM as well (postcoder), and just ran a usual optimizer with
2kbit permutation table and postcoder parameters, and results are
rather nice.
Of course its highly impractical to optimize the reordering for each block
or file, as it takes multiple hours, but there's not much sense in that
either probably, as it should be possible to precalculate a few
permutation tables beforehand, and then select one based on eg.
o1 frequency table analysis.