_R_CHECK_DEPENDS_ONLY_ vs. packages in .Library [was: Check package without suggests]

(Moving this one idea to R-devel)

? Wed, 19 Jul 2023 09:21:46 +0200
Henrik Bengtsson <henrik.bengtsson at gmail.com> ?????:

If you're on macOS, and have installed R the default way, it takes
more work to test on that platform. It works out of the box on Linux
and MS Windows.  See the '[R-SIG-Mac] CRAN installer for macOS -
directory permissions' thread started in April 2022
<https://stat.ethz.ch/pipermail/r-sig-mac/2022-April/014371.html>,
continued in May 2022
<https://stat.ethz.ch/pipermail/r-sig-mac/2022-May/thread.html>, and
June 2022
<https://stat.ethz.ch/pipermail/r-sig-mac/2022-June/014469.html>. It
was then renamed to 'System-wide site library [Was: CRAN installer
for macOS - directory permissions]' in June 2022
<https://stat.ethz.ch/pipermail/r-sig-mac/2022-June/014501.html>.

Currently, a check with _R_CHECK_DEPENDS_ONLY_=TRUE assumes that
.Library only has base and recommended packages. This assumption can be
broken on macOS, and also on other operating systems when R is
installed into a writeable directory or is running without installation
(e.g. R-devel from an SVN checkout) and the user doesn't pre-create a
separate library.

What would be the downsides to implementing _R_CHECK_DEPENDS_ONLY_ the
same way that _R_CHECK_NO_RECOMMENDED_ is already implemented? The
latter works by creating fake packages (with a DESCRIPTION and an empty
file called "dummy_for_check" but nothing else in them) in a temporary
library that take precedence over the ones in .Library and fail loading.

Best regards,
Ivan