You are absolutely correct that the difference between y against x and x against y fitting perfectly demonstrates why the bias exists, but the correct way to remove the bias is not normalization, but to use a coordinate-independent regression technique.
See the other comments by many other commenters for details.
I found it in the middle of teaching a stats class, and feel embarrassed.
I guess normalising is one way to remove the bias.