Fix stanr/stanli backend dropping group-level parameter levels

[gravesti]

Thanks for putting together this patch for brms. I came across this bug where the wrong variables were saved compared to cmdstan. LLM generated patch below.

read_stanr_fit_as_stanfit()/.stanfit_from_csfit() computed a correctly filtered 'variables'/'model_pars' list (excluding raw non-centered z_* parameters, keeping lprior/lp__), but never applied that filter to the actual 'samples' data frame before setting fnames_oi <- colnames(samples). This left fnames_oi out of sync with dims_oi/pars_oi, which downstream renaming logic assumes are in sync. Once a grouping factor had more than 2 flattened elements, this silently corrupted parameter names: raw z_1 columns were kept, lprior/lp__ were dropped, and one level of the derived r___ parameter went missing.

Fix: filter 'samples' down to columns matching model_pars/special_vars before computing fnames_oi, mirroring how read_csv_as_stanfit() already pre-filters columns via cmdstanr::read_cmdstan_csv(variables = ...).

Adds a regression test (tests/testthat/tests.stanr-backend.R) that fails on the prior code and passes after this fix.

Reprex:

## Minimal reproducible example for a data-loss bug in the 'stanr' backend
## (which also affects 'stanli', since stanli is embedded inside stanr).
##
## Bug: for a grouping factor with 3+ levels, brmsfit objects fitted via
## backend = "stanr" lose one level of the derived group-level effect
## (r_<group>__<term>[level,coef]) and the 'lprior'/'lp__' scalars, while
## retaining the raw, non-centered 'z_1' parameters that should have been
## dropped. A 'cmdstanr' (or 'rstan') fit of the exact same model is correct.
##
## Root cause (see brms/R/backends.R, read_stanr_fit_as_stanfit() /
## .stanfit_from_csfit()): the filtered 'variables' vector (which correctly
## excludes 'z_1' and keeps 'lprior'/'lp__') is used to build the metadata
## bookkeeping (model_pars/par_dims) but is never applied to filter the
## actual `samples` data frame, so `fnames_oi <- colnames(samples)` reflects
## the wrong (unfiltered) set of columns.

library(brms)

set.seed(6231)
n <- 60
dat <- data.frame(
  y = rnorm(n),
  x = rnorm(n),
  g = factor(sample(c("a", "b", "c"), n, replace = TRUE))
)

fit_common_args <- list(
  formula = y ~ x + (1 | g),
  data = dat,
  chains = 1,
  iter = 40,
  warmup = 20,
  seed = 8213,
  refresh = 0
)

fit_stanr <- do.call(brm, c(fit_common_args, list(backend = "stanr")))
fit_cmdstanr <- do.call(brm, c(fit_common_args, list(backend = "cmdstanr")))

vars_stanr <- variables(fit_stanr)
vars_cmdstanr <- variables(fit_cmdstanr)

cat("stanr variables:\n")
print(vars_stanr)
cat("\ncmdstanr variables:\n")
print(vars_cmdstanr)

r_g_stanr <- grep("^r_g", vars_stanr, value = TRUE)
r_g_cmdstanr <- grep("^r_g", vars_cmdstanr, value = TRUE)
z_leftover_stanr <- grep("^z_1", vars_stanr, value = TRUE)

cat("\nNumber of r_g_Intercept levels -- stanr:", length(r_g_stanr),
    " cmdstanr:", length(r_g_cmdstanr), " (expected: 3 for both)\n")
cat("Leftover raw z_1 parameters in stanr fit (expected: none):",
    length(z_leftover_stanr), "\n")

stopifnot(
  "stanr fit is missing group-level effect levels" =
    length(r_g_stanr) == length(r_g_cmdstanr),
  "stanr fit retains raw z_1 parameters that should have been excluded" =
    length(z_leftover_stanr) == 0
)