Fix Math-BigInt chained temporaries and arithmetic

[fglock]

Summary

• Fix localized global restore so foreach aliasing of localized $_ cannot corrupt method-chain temporaries.
• Invalidate bless-id overload classification with method/MRO cache invalidation so Math::BigInt::Calc picks up inherited overloads after @isa is populated.
• Add focused Math-BigInt regression coverage for bexp chaining, inherited overload classification, bmul, bmuladd, and bmodpow.

Tests

• timeout 1200 make > /tmp/make-math-bigint-pr.log 2>&1; printf EXIT line appended; EXIT: 0
• All 49 src/test/resources/module/Math-BigInt/t/*.t via timeout 300 jperl per file; all EXIT: 0

Notes

• Full bundled-module verification still has YAML-only failures in this workspace; YAML is being investigated separately in ../PerlOnJava3/.

[fglock]

Updated this PR to include PR #710 for combined testing.

Included commits:

• 38b9e6a fix: resolve bareword and special variable parsing issues in YAML/Mo.pm
• 936cffd merge: include YAML fixes from PR fix: resolve bareword and special variable parsing issues #710
• 8fe7d2d fix: keep qualified underscore identifiers packageable

Current verification before push:

• make passed: /tmp/make-pr709-with-yaml-2.log, EXIT: 0
• make test-bundled-modules is/was running in /tmp/test-bundled-modules-pr709-with-yaml.log; I will post the final result when it completes.

[fglock]

Combined verification result after including PR #710:

• make passed: /tmp/make-pr709-with-yaml-2.log, EXIT: 0
• make test-bundled-modules ran on commit 8fe7d2d: /tmp/test-bundled-modules-pr709-with-yaml.log, EXIT: 2
• YAML failures are gone.
• Math-BigInt failures are gone.
• Remaining bundled-module failure: module/Unicode-Collate/t/illegal.t, reported at line 61 as not ok 60.

Focused check:

• jperl direct run of Unicode-Collate/t/illegal.t produces not ok 60 and not ok 63 but exits 0: /tmp/unicode_collate_illegal_pr709.log
• system perl direct run passes all 176 assertions: /tmp/unicode_collate_illegal_system.log

[fglock]

Update after rebasing and force-with-lease pushing PR #709 to 1f4b7b9.

Included the PR #710 YAML commits on this branch:

• c9d1097 fix: resolve bareword and special variable parsing issues in YAML/Mo.pm
• dfaaa09 fix: keep qualified underscore identifiers packageable

Additional fixes in the latest commit:

• Preserve explicit surrogate Perl scalars so Unicode-Collate illegal.t no longer collapses adjacent surrogate scalars into one Java supplementary code point.
• Restore defined main-package ::sub calls while keeping missing ::bareword fallback, which removes the leading-:: 0/0 core-test regressions from the YAML merge.

Verification after rebase:

• make: PASS (/tmp/make-pr709-after-rebase-surrogate-leading-colon.log)
• Math-BigInt predicate: PASS (/tmp/math_bigint_predicate_after_rebase.log)
• module/Unicode-Collate/t/illegal.t: PASS (/tmp/unicode_collate_illegal_after_rebase.log)
• make test-bundled-modules: PASS (/tmp/test-bundled-modules-pr709-after-rebase-final.log)

Log compare for PR7027 vs PR709:

• The large reported regression is dominated by regex tests recorded as 0/0 in the external PR709 log. out.json shows ClassNotFoundException: org.perlonjava.app.cli.Main for those entries.
• A local regex slice on the rebased branch ran normally: re/reg_nocapture.t 25/25, re/qr.t 4/4, re/pos.t 7/8, so the regex 0/0 block appears to be an invalid external test run artifact.
• The remaining non-0/0 one-test drops reproduce at the PR709 after-counts locally: comp/package.t 12/14, japh/abigail.t 103/130, op/tr.t 279/318, uni/package.t 15/18. Direct logs are in /tmp/pr709-*-direct.log.

GitHub CI restarted from the new pushed head and is currently running.

[fglock]

Updated PR #709 with commit 671beeab2.

The four non-0/0 compare regressions are restored against the PR7027 baseline:

comp/package.t      13/14
japh/abigail.t     104/130
op/tr.t            280/318
uni/package.t       18/18

Additional verification:

make                                      PASS  (/tmp/make-pr709-core-deltas-fix3.log)
make test-bundled-modules                 PASS  (/tmp/test-bundled-modules-pr709-core-deltas-fix.log)
Unicode-Collate t/illegal.t               PASS  (/tmp/unicode_collate_illegal_pr709_core_deltas.log)
Math::BigInt bexp predicate               PASS  output: 2
CLI -s after -- predicate                 PASS  output: Just another Perl Hacker

The focused four-file rerun log is /tmp/pr709-four-deltas-after-fix1.log.

[fglock]

Updated PR #709 with the DBIx::Class leak-tracer reproducer and the tail-call alias cleanup fix.

Verification run locally:

• make -> passed (/tmp/make-dbic-final-unit.log, exit 0)
• Math::BigInt bexp predicate -> passed (2, exit 0)
• New unit unit/refcount/dbic_leak_tracer_compose_namespace.t is included in make and passes
• Existing tail-call cleanup units pass in make

I also reran DBIx::Class t/52leaks.t. The original detached result source failure is gone; the run now reaches ok 4 and auto-checks 5 weak refs cleanly, then exits 255 because the earlier nested txn exception resurfaces during later cleanup (/tmp/dbic_52leaks_after_reproducer_fix.log). That looks like a separate exception/eval cleanup issue from the detached source regression.