I guess this is only because -d 0 -e 2 -P 0 ends up using the ClampedGradient predictor anyway, but adding the -P 0 would make it try to do lz77 which is not very useful?

Exactly. File size was actually identical, suggesting it does the LZ77 but then doesn't save it. Likely because efforts 1-3 have some hardcoded parameters.

I also changed it so -P 0 will use RLE for effort 5 too, as RLE 4 to 5 have a large density change (11.5 bpp to 9.3 bpp), but LZ77 5 to 6 had almost none (7.24 bpp to 7.2 bpp). Means the effort levels have better tradeoffs.

Needs to be added to the merge queue, if you don't mind