Reported scores from optimising models with random effects way off

[lentinj]

See the hidden bit in #249

print(cbind(params.fixout$value, params.rndout$value))
print(c(
    fixout = attr(params.fixout, "summary")$score,
    # TODO: obj.fix$fn(obj.rnd$env$last.par.best) != attr(params.rndout, "summary")$score, even though newton trace reports the former?
    #       Some sublety post-inner-optimisation I've not figured out.
    rndout = attr(params.rndout, "summary")$score ))