lme() with known level-one variances

If I understand your request correctly, you want to use something like
"weights=varIdent(...)" as an argument to lme().  varIdent and the other
varFunc constructors have an argument "fixed" that allow you to specify
values for some or all of the coefficients of the variance function.
See ?varIdent.  The actual error variance will be varFunc() * sigma^2,
where sigma^2 is estimated.

That's the problem.

As happens in meta-analysis as well, the problem is to estimate a model
with a variance component fixed. Not fixed up to a scale parameter. Fixed.

In meta-analysis the model is that within each trial a treatment effect
parameter is constant, and as the trial is large the variance of the
estimated treatment effect is very accurately known conditional on the
true treatment effect for that trial. The unconditional variance is then
the known conditional variance plus an unknown variance.

It doesn't seem that lme() is designed for this, and last time I tried to
do it I gave up and changed the model more or less as you suggest.