It's fairly easy to fix the proximal problem identified in #857, but we run into trouble with profile() just a little farther down the line.

• devfun2() (redefine deviance function on the variance/covariance scale rather than the Cholesky scale) assumes that the model is using the default unstructured-covariance parameterization.
  • this could actually be an opportunity to refactor things a little bit; at present devfun2() uses some slightly clunky machinery to transform parameters from the var-cov scale back to the Cholesky parameterization, maybe there could be a generic for this which would then apply to unstructured as well as structured cov matrices?

We get warnings from devfun2, not an error — the error occurs farther downstream — but I think fixing devfun2 will probably resolve the problem.