I'm well familiar with 6502 shenanigans, but never really looked into 2600 programming before. Looking into that source file though, I got the impression that Atari mapped the I/O right into zero page? Kind of surprising, but I guess with the 6507 the effective address space was only 8K, so why not. Certainly if you're trying to get all your computing done in the blanking intervals, it'd help if banging the video chip could be done that much quicker as well.

comments: https://news.ycombinator.com/item?id=46925934

One of the authors, Dennis Debro, has done a bunch more of these: https://github.com/DNSDEBRO/Disassemblies