That's a slightly different issue: the issue with solving normal equations directly is that the condition number of X^\top X is the square of the condition number of X, so it's much more significant than lu-vs-inv. It can matter a great deal for ill-conditioned linear regression problems where some features resemble each other.