Maybe I am missing something, but this approach is doomed to fail / produce small gains.
A tar file is a concatenation of different 512 bytes of data. The first block contains the metadata of the file, and the next blocks contains the content.
Of those first 512 bytes of data, only the first 100 bytes contains the text.
Trying to play with the orders of files to make 100 bytes of data fit next to each other in block of 512 bytes, it does not seems a successful approach.
gzip use a 32kb windows, that can include up to 64 tar headers. At the very best we are saving 100bytes * 64 => 6400 bytes alright 7kb every 32kb. For an optimal compression of the (32-7)/32 => 78%.
The author says that the standard gzip compressed to 45768 / 337920 => 13%
I believe there are better approaches...