Fix R-CMD-check GHA workflow for R-devel: pin roxygen2 7.2.3 on all platforms + tryCatch fallback

[Copilot]

roxygen2 7.3.x crashes on setMethod() calls for primitive generics ([[, [, $, etc.) via a parser_setMethod bug — hyperSpec has 60+ such registrations. The previous workflow skipped installing 7.2.3 on R-devel, falling back to broken 7.3.x. The tryCatch fallback then attempted git checkout HEAD -- NAMESPACE man/ but both files were gitignored and never committed, leaving no NAMESPACE for R CMD check.

Changes

.github/workflows/R-CMD-check.yaml

• Remove if: matrix.config.r != 'devel' from the roxygen2 7.2.3 install step — installs 7.2.3 everywhere. Evidence from CI logs shows the SET_GROWABLE_BIT ABI break is resolved in current R-devel (7.3.x loads fine); 7.2.3 should load too and lacks the parser_setMethod crash.
• Wrap devtools::document() in tryCatch with informative diagnostics on git-restore failure (captures system2() output/status rather than silently discarding it):

  ok <- tryCatch({
    devtools::document()
    TRUE
  }, error = function(e) {
    message("devtools::document() failed: ", conditionMessage(e))
    message("Restoring NAMESPACE and man/ from committed versions...")
    out <- system2("git", c("checkout", "HEAD", "--", "NAMESPACE", "man/"),
                   stdout = TRUE, stderr = TRUE)
    status <- attr(out, "status")
    if (!is.null(status) && status != 0L)
      message("git restore failed (files may not be committed): ",
              paste(out, collapse = "\n"))
    FALSE
  })
  if (!ok) message("Using committed documentation.")

• Re-add Session info step (was accidentally dropped).
• Fix R-oldrel dependency condition: matrix.config.r == '3.6' (rgl pin only needed for R 3.6, not current oldrel).
• Add R-CMD-check log as step summary (find check -name '00check.log'), always runs.
• Update checkout@v3 → @v6, add permissions: contents: read, FORCE_JAVASCRIPT_ACTIONS_TO_NODE24: true.

.gitignore

• Remove NAMESPACE and man/*.Rd — these need to be committable so the tryCatch git-restore fallback can actually work for future ABI-break scenarios.