Make the RubyGems test suite green under RUBY_BOX=1 - #9826

Ruby::Box leaves $? uninitialized, so exit status checks read exit 0
regardless of the real status. Under RUBY_BOX=1 this made
Gem::Source::Git#rev_parse swallow git rev-parse failures, and made the
cargo availability pend in the test suite misjudge a broken toolchain as
usable. Process.last_status reports the real status in both modes.

Ruby::Box gives each box detached copies of the stdio globals, so
reassigning or reopening $stdout/$stderr cannot capture output written
by Kernel#warn, Kernel#puts or subprocesses, and $VERBOSE = nil cannot
silence category warnings in assert_ractor children. Pend the affected
tests until https://bugs.ruby-lang.org/issues/21867 is resolved.

Marshal in the main box cannot resolve Gem:: and other boxed constants,
so loading dumps of Gem::Specification, Gem::Version or Date raises
ArgumentError under RUBY_BOX=1. Pend the affected tests until
https://bugs.ruby-lang.org/issues/22090 is resolved.

The suite is green under RUBY_BOX=1 now, so run it on ruby-head as a
canary to catch regressions on either side early.

yamllint rejects the redundantly quoted matrix name. The cargo tests
require the installed gem with -r, which bypasses gem activation under
RUBY_BOX=1 because the command line require does not go through the
RubyGems Kernel#require override, so require inside -e instead.

Under RUBY_BOX every ruby subprocess spawned by the specs prints a
two-line experimental warning to stderr, breaking specs that assert
clean stderr. Strip those lines when the host ruby has boxes enabled,
instead of injecting -W:no-experimental into RUBYOPT, because several
specs assert the exact RUBYOPT propagated to subprocesses.

The special-characters and whitespace contexts stubbed File.expand_path
for every argument. Under RUBY_BOX=1 with the turbo_tests workers,
Bundler::Settings is still unloaded when these examples run, and reading
the repository .bundle/config triggers a lazy require of
rubygems/yaml_serializer. The default gem activation path in
Kernel#require then resolves paths through the stub and crashes with
Errno::ENOENT on a fake gemspec path. Stub only the expansion of
bundler/setup, which is all these examples need faked.

The example passes RUBYOPT=-rnative_child to the bundle install
subprocess, but under RUBY_BOX=1 require_libraries_in_main_box calls
rb_require_string directly and bypasses the RubyGems Kernel#require
override, so the gem is never activated and the require raises
LoadError. This is an unreported ruby-core bug related to [Bug #21760]
item 3, and the scenario itself is legitimate, so skip it until the
Ruby::Box require semantics are settled upstream.

The recursion is not in RbConfig.expand. Inside a box, defined?($gvar)
does not see assignments made in that box, so mkmf have_devel? never sees
its own memo and recurses until the stack is exhausted.

Both guards were written before the bugs had tickets. The $VERBOSE one
is now [Bug #22282] rather than a relative of [Bug #21867], and the
RUBYOPT -r one is [Bug #22295] rather than unreported.

Co-Authored-By: Claude Opus 5 <[email protected]>

The extension build note, the Process.last_status switch and the cargo
require were all written before their bugs had tickets. Each one now
names the issue that has to land before the workaround can go.

Co-Authored-By: Claude Opus 5 <[email protected]>

Both examples assert on stderr captured with capture_output, which comes
back empty under Ruby::Box. They landed with the ABI-scoped
specification support after this branch was written.

Co-Authored-By: Claude Opus 5 <[email protected]>

Ruby::Box leaves $? uninitialized, so the three probes read a successful
status after a failing command and carry an empty ref date into
Time.xmlschema, which raises instead of the intended error.

Co-Authored-By: Claude Opus 5 <[email protected]>