Fail CI on R CMD check NOTEs (needs shiny-workflows `error-on` input)

Follow-up to #184. Blocked on rstudio/shiny-workflows#60.

State today

R CMD check on pkg-r/ is clean — 0 errors | 0 warnings | 0 notes — and make r-check-package keeps it that way:

cd $(PATH_PKG_R) && Rscript -e "devtools::check(document = FALSE, error_on = \"note\")"

Verified both directions: a clean tree exits 0, and reintroducing a single shiny::: call fails with Error: R CMD check found NOTEs (exit 2).

The gap

That gate is local only. CI still passes on a NOTE.

.github/workflows/check-r.yaml delegates to rstudio/shiny-workflows/.github/workflows/R-CMD-check.yaml, which calls r-lib/actions/check-r-package@v2 without setting error-on. The action defaults to '"warning"', and the reusable workflow exposes no input to change it. So a NOTE introduced in a PR that nobody happens to run make r-check-package on lands on main silently.

Not workaroundable from here:

• extra-check-args is for R CMD check arguments, not rcmdcheck's error_on.
• The .github/shiny-workflows/check.{sh,R} hook runs as the step before check-r-package, so it can't influence that step's inputs.

The work

Once rstudio/shiny-workflows#60 ships the input, this is a two-line change:

  R-CMD-check:
    uses: rstudio/shiny-workflows/.github/workflows/R-CMD-check.yaml@v1
    with:
      working-directory: pkg-r
      error-on: '"note"'

Then confirm on a throwaway PR that a deliberately introduced NOTE actually turns CI red.

Watch out for

error_on = "note" can be flaky in CI even when the package is locally clean, because of environment NOTEs that never appear on a dev machine — most commonly:

• checking CRAN incoming feasibility (unstable network / dev version numbers)
• checking for future file timestamps (clock-check service unreachable)

Both are plausible on the 9-platform matrix. If they show up, prefer narrowing (e.g. apply error-on to a single representative platform, or set _R_CHECK_CRAN_INCOMING_=false) over reverting to "warning" — a check that never comes back clean is a check nobody reads, which is part of how the #182–#186 parity bugs survived.

Alternative if the upstream input stalls

Add a dedicated single-platform job here running rcmdcheck::rcmdcheck(error_on = "note") directly, alongside the existing matrix. More YAML to own, but unblocked and easy to scope to one runner.

Read .github/workflows/check-r.yaml and rstudio/shiny-workflows#60 first to confirm that the reusable workflow exposes the error-on input. Update the R-CMD-check invocation after the upstream change, then use a throwaway PR with a deliberate NOTE to verify that CI fails as intended.