Super-sampling 2x2 improved things, particularly around the ground-wall interface, where there is similar lighting. But edges that had moderate/strong lighting adjacent to zero lighting didn't improve that much, which breaks my idea of ignoring the normals. I also reworked the textures to remove hi-lighting at the edges, which I think helps a little as well, but I don't think that's a good thing to start doing, because really it's about contrast between textures, not the textures themselves, and I would then have to create many more textures to handle all the wall shapes if i was to darken around the edges, as I'm currently using a single texture for all walls. A lot of work compared to just employing MSAA. Performance is fine using 2x2 SSAA.
Super-sampling 4x4 further reduced the edge aliasing to almost being unnoticeable, but it produced severe texture aliasing, I assume because the linear sampler only samples from 2x2 pixels, not 4x4. I'll have to check this. That would explain why I've seen code for manual texture filtering. Also with 4x4, the 5 buffers (color,normal,position,depth,final-light-map) are 7680x4068 each, so 625MB. I don't think that's very good, and the performance is bad, with a noticeably low frame rate.
Next, onto MSAA.