Add spec for ruby2_keywords hash passed directly as positional argument

[jessecai33]

Adds a spec asserting that a Hash carrying the ruby2_keywords flag is not copied or unmarked when passed directly as an ordinary positional argument (i.e. not through a splat). It complements the existing "makes a copy and unmark the Hash when calling a method taking (arg)/(**kw)/(*args)" specs, which all cover the splat path — the direct positional path was previously unspecified.

Motivation

JRuby 9.4.15.0 shipped a regression (fixed in jruby/jruby#9517) where JIT-compiled code replaced a flagged hash with an unflagged copy on exactly this path. Because the behavior wasn't covered by ruby/spec, the regression reached a release and silently corrupted Rails ActiveJob argument serialization (ActiveJob::Arguments.serialize checks Hash.ruby2_keywords_hash? on a hash received as a positional parameter). The JRuby maintainers asked for the regression spec to be ported here.

The spec calls the target method repeatedly so implementations that compile methods lazily also exercise their compiled path — the JRuby regression only manifested in JIT-compiled code.

Verification

• CRuby 3.4.5: passes
• JRuby 9.4.15.0 (with the regression): fails
• JRuby 9.4 branch with Preserve ruby2_keywords hash passed as positional argument jruby/jruby#9517 applied: passes

🤖 Generated with Claude Code

[eregon]

Thanks it looks good, could you address those two things?

[eregon]

Suggested change

	obj.singleton_class.class_exec do
	def single(arg)
	arg
	end
	end
	def obj.single(arg)
	arg
	end

[jessecai33]

Applied in 132e4e1.

[eregon]

While I understand this makes it easier to catch the bug if it's in the JIT, one could argue many specs could do this and the test suite would take a while even if run in interpreter.
Since MSpec supports -R <repeats> I think that's better, it's already possible to run each example N times and that way to ensure it's JIT-compiled (possibly lowering the JIT threshold to compile faster).
That's notably what ZJIT does and it has proven to work well.

So could you remove the 100.times do?

[jessecai33]

Removed in 132e4e1. Verified the spec still catches the JRuby 9.4.15.0 regression without the loop: it fails there under -Xjit.threshold=0 / -X+C, and also on default settings with mspec -R 100. A plain single run on 9.4.15.0 passes since the interpreter path was not affected — relying on -R for JIT coverage as you describe seems right.