R's lm() function removes factor levels (aside from intercept estimator) without reporting error or warning

Question

I am attempting to fit a linear model, using forward stepwise ANOVA and AIC selection criterion, to a fairly large unbalanced data set, with interaction terms. There are 13,