model compiled with model methods fails to fit with future multisession

[n-kall]

Describe the bug

Trying to precompile model methods via compile_model_methods = TRUE (to be used after fitting) causes the fits to fail if using future.plan(multisession)

Error in .Call(<pointer: (nil)>, data_path, seed) : 
  NULL value passed as symbol address

To Reproduce

library(cmdstanr)

library(future)
library(future.apply)

plan(multisession)

m <- cmdstan_model(
  write_stan_file(
    paste0(capture.output(print_example_program()), collapse = "")
  ),
  compile_model_methods = TRUE
)

dat <- list(N = 3, K = 1, y = c(1, 0, 0), X = matrix(c(1, 1, 1), nrow = 3))

future_lapply(1:3, function(x) m$sample(data = dat), future.seed = TRUE)

Expected behavior
Sampling should work as expected, or if additional arguments are needed (e.g. for globals?) then guidance in documentation could help

Operating system
ubuntu 24.04

CmdStanR version number
0.9.0

[n-kall]

Doing some reading (e.g. https://future.futureverse.org/articles/future-4-non-exportable-objects.html), I have concluded that it is not possible to pass the model method functions (compiled/loaded via Rcpp) to child R processes. As this causes an error even though the model methods are not actually used in the code, this might be worth mentioning in the documentation for $compile and/or $init_model_methods?