[klauspost]

You would need to be able to reconstruct the input file bit-by-bit. S3 Clients expect to get back what they sent, exactly.

This puts a serious limitation on your compression. You would only be able to re-do the entropy coding part of DEFLATE, which is actually pretty good.

You would still need to store the original Huffman tables for each block, so you can reconstruct the entropy coding exactly.

I doubt this would even gain you a single percentage.