load_all() drops exports of generics re-exported from base

[SNAnalyst]

Summary

A package that provides S3 methods for a base generic can export the generic's
name so that pkg::plot(x) works without the package being attached. This is
export(plot) with no object named plot in the namespace; R resolves the
export through the namespace's parent chain, and pkg::plot becomes base::plot.

When the package is installed, this works exactly as intended. Under
load_all(), setup_ns_exports() removes such names from the package's exports,
so pkg::plot does not exist. load_all() and an installed package therefore
disagree about the package's API.

This is not the same as #31. I'm not arguing the warning is wrong — I understand
the position there that plot is genuinely "exported but not in the package
namespace". The point here is the second half: after warning, setup_ns_exports()
drops the export, and that is what makes load_all() diverge from an installed
package. A developer working under load_all() sees pkg::plot() fail with an
error that does not occur for their users.

Reprex

No roxygen2 involved; hand-written NAMESPACE.

basegen/
  DESCRIPTION      # Package: basegen, plus the usual required fields
  NAMESPACE
  R/plot.R

# NAMESPACE
S3method(plot,widget)
export(plot)

# R/plot.R
plot.widget <- function(x, ...) invisible("widget method ran")

Installed, this behaves as intended:

"plot" %in% getNamespaceExports("basegen")
#> TRUE
identical(basegen::plot, base::plot)
#> TRUE
basegen::plot(structure(1, class = "widget"))
#> "widget method ran"

Under load_all():

pkgload::load_all("basegen", export_all = FALSE)
#> Warning: Objects listed as exports, but not present in namespace:
#> • plot

"plot" %in% getNamespaceExports("basegen")
#> FALSE

basegen::plot(structure(1, class = "widget"))
#> Error: 'plot' is not an exported object from 'namespace:basegen'

Mechanism

setup_ns_exports() (R/namespace-env.R, current dev) keeps only names it finds
in the namespace env or the imports env, and does not follow the parent chain to
base:

ns_and_imports <- c(env_names(nsenv), env_names(imports_env(package)))
extra_exports  <- setdiff(exports, ns_and_imports)
...
exports <- intersect(ns_and_imports, exports)   # base-inherited exports dropped here

Consistently, a re-export from a normal package is fine, because the object
lands in the imports env:

# NAMESPACE:  export(head)  +  importFrom(utils, head)
pkgload::load_all("basegen", export_all = FALSE)
"head" %in% getNamespaceExports("basegen")   #> TRUE
"plot" %in% getNamespaceExports("basegen")   #> FALSE

So base is the special case: the object it re-exports is reachable through the
parent chain but is neither in the namespace env nor the imports env.

No workaround

The obvious one, importing from base, is rejected by R itself:

# NAMESPACE:  export(plot)  +  importFrom(base, plot)
pkgload::load_all("basegen")
#> Error: operation not allowed on base namespace
install.packages("basegen", repos = NULL, type = "source")
#> Error in asNamespace(ns, base.OK = FALSE) : operation not allowed on base namespace

@rawNamespace export(plot) produces the same NAMESPACE line and the same
outcome. So a package in this position cannot make load_all() match an
installed build.

Suggestion

Would setup_ns_exports() be willing to keep exports that resolve in the
namespace's parent chain — e.g. exists(name, envir = nsenv, inherits = TRUE),
or at minimum names that resolve in baseenv() — rather than dropping them? That
would make load_all() agree with an installed package for re-exported base
generics, independently of whether the warning is kept.

Happy to send a PR if that direction is acceptable.

Context

This bit us in practice: working under load_all(), a contributor saw
pkg::plot() fail and removed the @export tag to "fix" it, which then broke the
installed package for every user. Reduced example above.

pkgload 1.5.3 (current CRAN); dev setup_ns_exports() is unchanged. R 4.6.1.