Ark: Should we ensure subprocesses run the started version of R?

(At first I thought yes, but I ended up convincing myself that the current behaviour is best.)

Positron has the ability to start a version of R that's not the default on the PATH (RStudio only offers that feature on Windows). However subprocesses created by calling back into R will still reach into the default R version reachable from PATH.

To reproduce, start a version of R that's not the default, for example with 4.1.2 running and 4.3.1 default:

system("R --version")
#> WARNING: ignoring environment value of R_HOME
#> R version 4.3.1 (2023-06-16) ...

I thought we could add $R_HOME/bin to PATH to ensure these calls hit the correct R version:

modified   crates/ark/src/interface.rs
@@ -179,6 +179,11 @@ pub fn start_r(
     let libraries = RLibraries::from_r_home_path(&r_home);
     libraries.initialize_pre_setup_r();
 
+    let r_home_path = r_home.display();
+    let path = std::env::var("PATH").unwrap();
+    let path = format!("{r_home_path}/bin:{path}");
+    std::env::set_var("PATH", path);
+
     crate::sys::interface::setup_r(args);
 
     libraries.initialize_post_setup_r();

But after more research I don't think it's needed because it's deviating from normal R behaviour. First note the warning about ignored R_HOME above. It's from the R wrapper script. See R.sh.in (with R_HOME_DIR hardcoded at build time):

if test -n "${R_HOME}" && \
   test "${R_HOME}" != "${R_HOME_DIR}"; then
  echo "WARNING: ignoring environment value of R_HOME"
fi

The wrapper script could start the R from R_HOME but chooses not to and warn instead.

Furthermore the R subprocess calls are careful to pick the R_HOME version of R. For instance in install.packages():

cmd0 <- file.path(R.home("bin"), "R")
args0 <- c("CMD", "INSTALL")

Same for callr which starts the correct R version:

callr::r(function() getRversion())
#> [1] ‘4.1.2’

Finally, rig doesn't do anything to correct that either. Even after running rig system make-orthogonal, changing the default R version while running R will cause new calls to system("R") to pick up the new default instead of the started version. (cc @gaborcsardi)

So on second thought I don't think we need to do better, it's preferable to stick with the regular behaviour to prevent surprises when running code in a different setup. WDY'allT?

Read crates/ark/src/interface.rs and the referenced R wrapper script R.sh.in; compare the reproduction with R's R.home handling and the behavior of install.packages() and callr. Done means deciding and documenting whether Ark should alter PATH, including whether regular R behavior should be retained.