I doubt I'll be able to fix this kind of aliasing because of the Nyquist limit; something with jagged edges should be easier (you had an example in another thread).
Also, I tried to find out why edge-directed interpolation worked with dual ISO, and it seems to be because lines are grouped in pairs (2 scanned, 2 skipped), unlike here (one line sampled, a bunch of them skipped). That means, with dual ISO I have high-frequency detail enough to estimate diagonal lines and fine detail like hair or fur, but here I don't have any.