Modernize registration of S3 method for projpred generic

[fweber144]

I'm not sure if you are willing to increase the required R version to >= 3.6.0 (until now: >= 3.5.0), but if you are, then this PR provides a more modern way to register the S3 method get_refmodel.brmsfit() for the projpred::get_refmodel() generic.

This was already discussed in #1222, but at that time (more precisely, at #1222 (comment)), I didn't realize that you don't need to call library(projpred) before documenting via devtools::document() if you can live with having get_refmodel.brmsfit( in man/get_refmodel.brmsfit.Rd (as is the case already now) instead of \method{get_refmodel}{brmsfit}( (the latter would render the typical ## S3 method for class 'brmsfit' header in the "Usage" section of HTML and PDF documentation, but as I said, we don't have that header at the moment either). In other words, in #1222, I didn't realize that calling library(projpred) before documenting only provides a minor cosmetic improvement compared to the current state.

Another reason why I would switch to this more modern S3 registration is that in https://github.com/r-lib/vctrs/blob/master/R/register-s3.R, the code for s3_register() has changed, so it should probably be updated if you don't want to adopt the more modern scheme. However, when copying the new vctrs::s3_register() code, you have to keep in mind that this would require at least one new helper function (vctrs:::.rlang_s3_register_compat()), which increases the amount of code that has to be copied to and maintained in brms. This also shows that keeping the current S3 registration up to date is kind of tedious, because you regularly have to check vctrs::s3_register() for updates.

I have tested this PR locally with the following code (where I restarted the R session between each section):

# Attach ------------------------------------------------------------------

library(brms)
options(mc.cores = parallel::detectCores())
ref_fit <- brm(
  formula = rate ~ conc + state,
  data = Puromycin,
  backend = "cmdstanr",
  seed = 2052109,
  refresh = 0
)

library(projpred)
prj <- project(ref_fit, solution_terms = "conc", ndraws = 40)
prjmat <- as.matrix(prj)

# Load --------------------------------------------------------------------

options(mc.cores = parallel::detectCores())
ref_fit <- brms::brm(
  formula = rate ~ conc + state,
  data = Puromycin,
  backend = "cmdstanr",
  seed = 2052109,
  refresh = 0
)

prj <- projpred::project(ref_fit, solution_terms = "conc", ndraws = 40)
prjmat <- as.matrix(prj)

# Attach brms, load projpred ----------------------------------------------

library(brms)
options(mc.cores = parallel::detectCores())
ref_fit <- brm(
  formula = rate ~ conc + state,
  data = Puromycin,
  backend = "cmdstanr",
  seed = 2052109,
  refresh = 0
)

prj <- projpred::project(ref_fit, solution_terms = "conc", ndraws = 40)
prjmat <- as.matrix(prj)

# Load brms, attach projpred ----------------------------------------------

options(mc.cores = parallel::detectCores())
ref_fit <- brms::brm(
  formula = rate ~ conc + state,
  data = Puromycin,
  backend = "cmdstanr",
  seed = 2052109,
  refresh = 0
)

library(projpred)
prj <- project(ref_fit, solution_terms = "conc", ndraws = 40)
prjmat <- as.matrix(prj)

as well as with a couple of hands-on examples as well as within projpred's unit tests (and I have also run R CMD check for brms, of course).

[fweber144]

If you are also willing to always run library(projpred) before documenting, then I could push a commit that updates the documentation correspondingly (i.e., which provides the cosmetic improvement mentioned above). But in #1222, I understood you in the way that this complicates your build toolchain, so I sticked with the "standard" way of documenting (i.e., without calling library(projpred) beforehand).

[paul-buerkner]

Thank you! Yes, I would prefer not to load projpred before documenting. I will check out the PR now.

[paul-buerkner]

Looks good. Merging now.