ci: restore R.utils, which the dependency trim removed by mistake

R.utils is named in no chapter, which is why b4d3666 dropped it as historical
leftovers. It is needed anyway: yulab.utils' download_yulab_file() calls
R.utils::gunzip() when gzfile = TRUE, and the only copy of HDO.sqlite on the
server is the gzipped one, so gzfile = TRUE is unavoidable here.

The failure is silent by design. check_pkg() aborts with a missing_package_error,
and download_yulab_file wraps the download and the gunzip in one tryCatch that
converts any error to FALSE -- it then walks to the next URL, fails the same way,
and returns the destination path as though it had succeeded. The caller sees a
plausible path, checks file.exists(), and concludes the network was unreachable.
That is why the previous commit's retry() wrapper could not help: the download
was never the thing failing.

The timeline matches. The last successful publish, on 2026-10-11, ran with
R.utils installed and passed this step. Every run since b4d3666 landed has
failed here, which is why the site has been frozen at the 10-11 build.

Proved by running the step's own R payload out of the workflow file, against a
fresh HOME, twice:

  library without R.utils -> exit 1; the download completes to 100% and leaves
                             HDO.sqlite.gz behind with no HDO.sqlite
  library with R.utils    -> exit 0; HDO.sqlite written, 10.95 MB

The lesson is about the method rather than the package. That trim was decided by
searching the .qmd sources for mentions, which cannot see a dependency that
lives inside another package's code. Same reason the same commit had to keep
ggupset and ggridges, which no chapter names either.

Author: GuangchuangYu

.github/workflows/publish-book.yml

@@ -78,6 +78,16 @@ jobs:
           # executed chunks. They currently arrive only as transitive dependencies of
           # clusterProfiler, enrichplot and biomaRt; a chapter that calls a package
           # should not depend on another package happening to import it.
+          #
+          # R.utils looks unused -- no chapter mentions it -- but yulab.utils needs
+          # it: download_yulab_file(..., gzfile = TRUE) calls R.utils::gunzip()
+          # after fetching the ontology cache, and the only copy of HDO.sqlite on
+          # the server is the gzipped one. check_pkg() then aborts, and
+          # download_yulab_file swallows that into FALSE and returns the
+          # destination path as though the download had worked, so the cache is
+          # silently absent and the HDO prewarm step fails. Do not remove it on
+          # the grounds that it is unreferenced in the .qmd sources: trimming by
+          # text search cannot see a dependency that lives inside another package.
           retry Rscript -e 'options(Ncpus = 1); pak::pkg_install(c(
             "cran::dplyr",
             "cran::ggplot2",
@@ -94,6 +104,7 @@ jobs:
             "cran::wget",
             "cran::httr2",
             "cran::dotenv",
+            "cran::R.utils",
             "cran::ggupset",
             "cran::ggridges",
             "cran::shadowtext",