Yes, the baseline model here is identified when the mean is constrained, but the standard errors are still not correct (they don't take into account the equality information in heq, so some may be to large/small).
There are really only two solutions in my mind to fix this: 1) some transformation matrix is built up based on the heq() function so that pre-post multiplication of the unconstrained hessian is possible, or 2) don't use the EM, and instead adapt the BL method to compute this matrix numerically. The latter seems more feasible at the moment because conceptually the former is harder to obtain (there could be a trick that I'm not seeing though). Either way, I think for now I'll have to disable the SE computations when the alabama or Rsolnp algorithms are used.