It is a matter of the degrees of freedom left for the error sum of squares. I suspect that if you check the Analysis of Variance outline in the Fit Least Squares platform, you will find that the DF for Error is 1. That constrains the residuals to be two levels. There is nothing wrong, you just don't have the independent information to really see the model errors. (Residuals from the model estimate the errors from the response.) If you remove higher order term, you will see the number of unique residuals levels double from 2 to 4.