I’ll do my best to answer your question with what I know.
I think that deriving the normal from two separate alpha channels is actually higher quality (though you can’t perceive it) than the regular individual normal map texture compression that UE uses. But I think doing that adds a greater memory footprint than using 3 x RGB only textures (including an individual normal map).
However, the more important thing, I didn’t really dwell on before, is that at higher resolutions (4k+) the compression artifacts that you’re seeing in your demo there are waaaay less noticeable. As I said, I had to bring the resolution down to 512x512 before I could really begin to notice the compression artifacts messing with the overall appearance.
So my conclusion is that texture packing into 2 x RGB only textures (and deriving normals from the two G channels) where possible when dealing with 4k textures is a big memory and lookup saver, without compromising too much on overall appearance. At lower resolutions those savings aren’t worth it because of the visual fidelity loss.
What would be good in future would be an image standard that allowed unlimited channels and the ability to choose how you compress each one. Just a thought Epic ; )