See stan-users discussion.
I think the sample file that's output from variational inference
should match the one output for MCMC if at all possible.
Would it be possible to take that summary of means and write it
out as a comment? Or in a separate output file?
Going forward, we're going to split the outputs in Stan 3 (if
I get my way) so that we don't try to overload CSV notation with
out-of-encoding comment information that needs to be parsed, which
is too hard to deal with other than with our own I/O tools (which
themselves become hard to maintain).
See stan-users discussion.