Actually, I think it makes sense now. When you do a read, it reads a combination of the la_data_in and the current value that was written depending on the la_oen bits. I think it still would've made more sense to have dedicated input and output mapped addresses rather than a mix, but shrug

This is an artifact of me looking at the code and thinking that it was needlessly preventing full duplex on all 128 channels and fixing that in the most expedient way possible. All blame can go to me if it seems quirky.