Match absolute paths in the irb_require already-loaded check

[youdie006]

#1253 anchored the already-loaded check with \A, but $LOADED_FEATURES holds absolute paths for anything the real require loaded, so the guard can no longer match them:

require 'json'          # $" gets "/usr/lib/ruby/3.2.0/json.rb"
irb_require 'json'      # the guard misses, so this loads and evaluates it again

Measured against the shapes that actually appear in $":

file_name	$" entry	before #1253	on master	with this change
foo	/tmp/x/foo.rb	matches	misses	matches
foo	foo.rb (what irb_require itself pushes)	matches	matches	matches
foo	/tmp/x/foo2.rb	matches (the #1252 bug)	misses	misses
foo	/tmp/x/barfoo.rb	matches (same bug)	misses	misses

Anchoring on a path boundary rather than the string start keeps both: absolute entries match again, and the foo/foo2 substring match #1253 was fixing stays rejected.

Verification

Base d05df55, Docker ruby:3.4.

row	load.rb md5	new test	#1253's test_irb_require_file_matches_exact_name
master	3b81c5e261	1 failure	passes
this PR	f15815509a	passes	passes
drop the trailing \z	ec0bb20c9e	1 failure	passes
drop the boundary entirely	4f7a97b201	1 failure	passes

The two mutants are killed by different rows of the new test: the \z by the irb_require 'irb_require' case, the boundary by the irb_require 'probe' case.

Full suite with the fix: 409 tests, 2496 assertions, 1 failure. That one is test_use_tracer_enabled_when_gem_is_available, which also fails on unmodified d05df55 with the unmodified test file (408 tests, 1 failure) because the tracer gem is not installed in the container; load.rb does not mention tracer. rubocop: 116 files inspected, no offenses.

Not run here: rake test_in_isolation, the yamatanooroti integration tests (need libvterm, no sudo), the debug-compat job, and the Ruby 2.7 to 3.3 and prism 1.3.0/1.7.0 matrix rows. The change is a regex literal, so cross-version risk is low, but I have not verified it.

One observation while writing the test, offered as something worth extending rather than a complaint. test_irb_require_file_matches_exact_name still passes with the complete pre-#1253 regex restored, so it does not reproduce #1252: it loads foo then foo2, and foo is not a substring of foo.rb at that point. Reproducing #1252 needs foo2 loaded first, since foo is a substring of foo2.rb.

I also left irb_require pushing relative names into $" at lib/irb/command/load.rb:55 and :66 alone, and did not widen the extension list beyond the existing (?:\.o|\.rb)?, so a .so feature still will not match. Both seemed like separate scope.

---

Disclosure: written with the help of Claude (an AI assistant). Every number above is from runs on my machine.

[youdie006]

Pushed a fix for the red rows, which were mine: the probe file read $irb_require_probe_count before it was assigned, so under -w Ruby warns on stderr and the test's own assert_empty(err) caught it. The counter is now initialised in the session first.

My local run had not enabled warnings, which is why I missed it; re-ran the whole matrix with RUBYOPT=-w this time.

row	new test
master	1 failure
this PR	passes
drop the trailing \z	1 failure
drop the boundary	1 failure

Full suite under -w: 409 tests, 2496 assertions, 1 failure, which is the pre-existing test_use_tracer_enabled_when_gem_is_available (the tracer gem is absent in my container; unmodified d05df55 gives 408 tests and the same single failure).