Okay I have a 52-frame improvement to replace this movie with at:
http://tasvideos.org/userfiles/info/69339170071981656
RNG was unfortunately slightly worse in all four stages this time around, and I forgot to mention this earlier that there is a strange framerule-like system that the game uses which limits the number of suitable RNG patterns to test. For example, in the case of the fourth stage, this was the only RNG pattern which made the fastest framerule.
[quote nymx]...but I'm pretty sure that a game like this doesn't make use of subpixels.[/quote]
It appears you are correct! I was using a RAM address which I thought tracked X subpixels but upon closer inspection, it does not. Visually, it is actually possible to see that the manipulation I make allows me to jump from the very edge pixel of the platform, whereas without the manipulation the rightmost I can jump is two pixels to the left of that.
I have already updated the TAS notes and temporary encode to reflect all of the above.