[avehtari]

In get_refmodel.stanreg(), cvrefbuilder passes dis = dis, and the function sets dis to the full-data sigma draws only after defining cvrefbuilder. R evaluates dis when cvrefbuilder runs, so in cv_varsel(cv_method = "kfold") every fold-wise reference model of a Gaussian model used the full-data sigma instead of the sigma of its own refit. The fix passes the user-supplied dis (usually NULL), so each fold takes sigma from its refit. The new test in test_refmodel.R fails without the fix.

Effect, measured on Gaussian stan_glm() models with 10-fold CV (N 30 to 400, P 5 and 25, 100 simulated data sets per cell): the old code overestimated the reference model's elpd by about 1 for every N. For N ≥ 100, it overestimated the submodels' elpd by 0.15 to 1, growing with the submodel size; that is at most 0.15 SE on average, and up to 0.7 SE for N 30. The rankings do not change, but the suggested size changed in up to 13 % of the data sets.

brms has the same pattern in get_refmodel.brmsfit(), which needs a separate PR.

This PR was assisted by Claude