>>18ah, so you want to dynamically look for the best possible transform, then embed metadata about which transform was used, and then potentially make more rounds of searching, applying and embedding as long as the size of output is lower than the size of input? then, to decompress you would do no searches but just read what transform was used from metadata, apply the reverse transform, then read extracted metadata and repeat the process until no compression metadata is present?
this could potentially work (although it would probably be extremely inefficient when it comes to compression time, and relatively inefficient with decompression), although the pigeonhole still applies.