Fix IO::Async B::SV refcount regression masked by suite timeout

[fglock]

Summary

IO::Async 0.805 regressed from PASS to FAIL in CPAN random-tester run 20260826-133506-83222. The report is recorded as a 300-second timeout, but the suite had already completed several failing tests before the timeout. The failures are reproducible reference-count mismatches in both PerlOnJava backends.

There are two related compatibility problems:

1. B::SV::REFCNT reports an extra owner for otherwise singly-owned IO::Async::Loop objects (expected 1, found 2).
2. The PerlOnJava patch for t/05notifier-loop.t still expects a historical extra notifier owner (3), while current PerlOnJava reports 2, which now matches system Perl.

The overall 300-second cap is a secondary concern: IO::Async deliberately contains multi-second timer tests, and the archived run was still making progress. This is not evidence of a hang.

Distribution and configuration

• Distribution: PEVANS/IO-Async-0.805.tar.gz
• Previous result: PASS on 2026-05-12 at PerlOnJava commit b5c01ce1c
• Current result: REGRESS on 2026-08-26
• Distropref patches applied:
  • IO-Async/NoFork.patch
  • IO-Async/PerlOnJava.patch
  • IO-Async/SkipUnsupportedSocketTests.patch
• Distropref environment: IO_ASYNC_NO_FORK=1

Configuration and build both succeed. The failures occur during Build test.

Focused results

System Perl

Using the same IO::Async 0.805 source and required test dependencies:

• t/05notifier-loop.t: PASS, 35/35
• t/11loop-poll-timer.t: PASS, 19/19

The final native-Perl expectation in t/05notifier-loop.t is 2 references.

PerlOnJava JVM backend

t/05notifier-loop.t:

not ok 35 - $loop has PerlOnJava notifier owner finally
# expected 3 references, found 2

t/11loop-poll-timer.t:

not ok 3 - $loop has refcount 1
# expected 1 references, found 2
not ok 5 - $loop has refcount 1 after watch_time
# expected 1 references, found 2
not ok 19 - $loop has refcount 1 finally
# expected 1 references, found 2

PerlOnJava interpreter backend

Both focused tests produce the same failures and counts as the JVM backend. This points to shared runtime/refcount behavior rather than backend-specific compilation.

Other failures visible before timeout

The archived CPAN run also reports expected 1, found 2 in both Poll and Select variants of:

• timer tests
• signal tests
• idle tests
• control tests
• metrics tests

The run reached tests after t/22timer-absolute.t before the tester's 300-second total-runtime limit fired. Therefore the timeout classification masks real completed failures.

Likely regression area

History and blame point to commit fe9d6bc5f (feat: complete Perl ithread lifecycle compatibility, 2026-08-15). That commit changed B::SV::REFCNT to call:

Internals::SvREFCNT($_[0]->{ref}, 1)

and added a special case in Internals.svRefcount() that increments the reported count when:

rc == 2
    && the B-specific flag is true
    && !ReachabilityWalker.hasLiveStrongScalarReferentOtherThan(base, arg)

The special case was introduced to keep a real aggregate owner visible in DBIx::Class/thread lifecycle paths. It appears to misclassify the temporary/private owner created by B::SV for these IO::Async loop objects, producing 2 where native Perl and IO::Async expect 1.

This regression point is inferred from the focused behavior and source history; it has not yet been confirmed by a commit-by-commit bisect.

The opposite-looking t/05notifier-loop.t result is a stale compatibility expectation: PerlOnJava now reports the native-Perl value of 2, but SkipUnsupportedSocketTests.patch changes that assertion to expect 3.

Proposed work

1. Add a small project-owned regression test that reproduces the false extra B::SV::REFCNT owner without requiring the full IO::Async suite. Validate it on system Perl first and on both PerlOnJava backends.
2. Refine the B-specific Internals::SvREFCNT(..., 1) heuristic so singly-owned IO::Async-style objects report 1 without breaking the DBIx::Class and thread lifecycle cases it was added to protect.
3. Update IO-Async-0.805/SkipUnsupportedSocketTests.patch so the final notifier-loop assertion expects the upstream/native count of 2.
4. Rerun the focused Poll and Select timer, signal, idle, control, and metrics tests on both backends.
5. Rerun the full IO::Async suite. Consider adding TEST_QUICK_TIMERS=1 to the IO::Async distropref, or a justified runtime exception, so the random tester can finish within its global runtime policy. Keep this timeout adjustment separate from the correctness fix.

Acceptance criteria

• The new focused refcount regression passes under system Perl, JVM PerlOnJava, and interpreter PerlOnJava.
• t/05notifier-loop.t and the affected Poll/Select loop tests pass on both PerlOnJava backends.
• Existing DBIx::Class and thread lifecycle/refcount regression coverage remains green.
• A bounded full jcpan -t IO::Async run completes and passes with the maintained distropref configuration.