The repetitions would have to appear within 32KB of each other to be captured in the DEFLATE window [1] which, given that a lot of pages can be 100s of KBs of uncompressed text, may make that impossible.
Additionally, the zlib (or whatever) compression settings would have to be set aggressively enough to identify those duplicates as the best run-lengths to encode. Since GitHub is generating a lot of this live and delivering with low TTFB, they may be using a less aggressive than necessary setting.
The repetitions would have to appear within 32KB of each other to be captured in the DEFLATE window [1] which, given that a lot of pages can be 100s of KBs of uncompressed text, may make that impossible.
Additionally, the zlib (or whatever) compression settings would have to be set aggressively enough to identify those duplicates as the best run-lengths to encode. Since GitHub is generating a lot of this live and delivering with low TTFB, they may be using a less aggressive than necessary setting.
[1] https://en.wikipedia.org/wiki/DEFLATE#Duplicate_string_elimi...