perf: reduce reference-counting overhead and DBIx::Class warnings

[fglock]

Summary

• register tracked scalar owners once after DESTROY-rescue handling instead of twice
• remove redundant synchronized wrappers from the explicitly single-threaded scalar registry
• remove empty reverse weak-reference buckets during unweaken
• stop caught optional base::import failures from printing duplicate stderr diagnostics
• align bundled DynaLoader and bytes file versions with their Java runtime versions
• patch Exception::Class 1.45 generated subclasses to match the source file registered in %INC
• preserve weak references to readonly constants while their defining subroutine remains installed
• add focused performance and version-consistency regression coverage

Performance

Three fresh JVM runs of dev/bench/benchmark_refcount_store.pl:

Revision	Median CPU time
master (168466619)	2.71 s
this branch	2.41 s

This is an 11.1% reduction in CPU time for tracked reference stores after weaken() activates lifecycle bookkeeping.

The ordinary method benchmark remained approximately 2.35 CPU seconds on both revisions.

DBIx::Class warnings

t/00describe_environment.t no longer reports:

• empty Getopt::Long::Descriptive base packages from caught optional-dependency probes
• Params::ValidationCompiler generated exception version mismatches
• DynaLoader or bytes runtime/file version mismatches

Exception::Class 1.45 still contains the upstream 1.1/1.45 mismatch; the CPAN distroprefs patch keeps the generated class version consistent with Exception/Class.pm while preserving its %INC registration.

Moo weak accessors

Readonly scalars created by \ "literal" are strongly owned by their defining subroutine's pad, but that edge is invisible to selective refcounting. Moo passes these references through constructor temporaries before weakening an attribute, which previously let mortal processing clear the value immediately.

The runtime now recognizes readonly scalars in installed subroutine pads as strongly owned. Replacing the subroutine still clears the weak reference through the existing optree-reaping hook.

Validation

• standard Perl: prove src/test/resources/unit/core_module_file_versions.t
• PerlOnJava: ./jperl src/test/resources/unit/core_module_file_versions.t
• Exception::Class 1.45: 7 files, 86 tests, all successful
• DBIx::Class t/00describe_environment.t: passed with none of the targeted warnings
• ./jcpan -t Moo: 71 files, 841 tests, all successful
• make: successful
• ./jcpan --jobs 8 -t DBIx::Class
  • 314 files
  • 13,121 tests
  • 695 seconds wall time after the Moo fix
  • all tests successful

Safety notes

Two broader reference-counting optimizations were evaluated and rejected after unit-test regressions:

• conditionally skipping temporary reachability roots
• resetting the global weak-reference mode after the last weak reference cleared

The retained performance changes preserve the existing lifecycle state machine. The warning changes retain catchable base exceptions and validate file/runtime version consistency with a portable test that also passes on standard Perl.

Generated with Codex