Moderation Model - negative estimates

I don't think you should try to interpret the results of the first model that lacks the main effect of age on the mean (and which returned Status Red). One way to think about this kind of biometric moderation model is that the phenotype is being regressed onto latent A, latent C, latent E, an interaction between age and latent A, an interaction between age and latent C, and an interaction between age and latent E. Not including the main effect of age is wrong in the same way that it's wrong in ordinary multiple regression to include an interaction term without including the main effects of the two predictors.

While it's true you obtained some negative point estimates from the second model, that's not necessarily bad. Your parameters a11, c11, and e11 represent what the loading (path coefficient) onto (respectively) A, C, and E will be when age equals zero. Your parameters aM11, cM11, and eM11 represent how much their corresponding loading changes for a 1-unit change in age. Note that the resulting loadings themselves are ambiguous with respect to sign. For instance, the A variance component is (a+ D%%aM) %% t(a+ D%*%aM), which in this case is equivalent to squaring a scalar, so it will always be non-negative.