docs(D-W6): empirical findings — drift source at sub install path

[fglock]

Status: empirical D-W6 progress, walker gate restored

This PR began as the full D-W6 implementation (replicate Perl 5 destroy
semantics exactly by removing the walker gate). The experiment was run,
documented, and partially reverted. Net change vs master is small:

• dev/modules/moose_support.md — full D-W6 plan + empirical findings.
• DestroyDispatch.java — new PJ_DESTROY_TRACE=1 debug flag for
  finding refCount drift sources. No behaviour change otherwise.
• MortalList.flush and RuntimeScalar.set — comments updated to
  point at the D-W6 plan; walker gate retained as a safety net
  (matches PR fix(walker-gate): replace class-name heuristic with universal walker check #599's universal walker).

What we learned

What removing the gate fixes (commit 230a672dd, then reverted)

Test	With gate (master)	No gate
t/cdbi/04-lazy.t test 11	FAIL	PASS
t/storage/txn_scope_guard.t 18/18	FAIL	PASS
t/52leaks.t 11/11	bailed	PASS

Cycles also leak naturally (verified with a small reproducer). Exactly
the Perl 5 behaviour the user asked for.

What removing the gate breaks

Moose bootstrap — use Class::MOP::Class dies with

Can't locate object method "initialize" via package "Class::MOP::Class"
  at jar:PERL5LIB/Class/MOP/Mixin.pm line 12.

PJ_DESTROY_TRACE=1 shows ~15 non-blessed RuntimeCode objects being
destroyed inside MortalList.flush during MiniTrait::apply —
while sub-installation is in progress. By the time the bootstrap
calls Class::MOP::Class->initialize(...), the initialize slot is
gone from the package stash.

Drift source identified

Anonymous CVs (sub { ... }) transiently hit refCount=0 between
creation and being installed in the package stash. The walker gate
absorbs this drift; without it, DESTROY fires on subs that should
still be live.

Hybrid (only destroy blessed) doesn't work

Skipping callDestroy for non-blessed objects breaks two things:

• cmop/numeric_defaults.t (12 fails) — anonymous metaclasses ARE
  blessed, and their refCount drift hits them.
• t/52leaks.t (4 fails) — callDestroy for non-blessed containers
  does cascade-decrement of contained blessed children; skipping it
  leaks DBIC rows transitively.

Next concrete steps (multi-week)

Each is a separate audit + reproducer + fix, documented in the design
doc:

1. Sub installation drift — BytecodeCompiler.compileNamedSub and
   GlobalVariable.defineGlobalCodeRef. The sub's first refCount
   increment must happen before any intermediate scalar in the
   bytecode emit chain drops its transient hold.
2. Closure capture drift — RuntimeCode.cloneForClosure and the
   addToCapturedScalars paths.
3. Argument promotion drift — RuntimeArray.push / @_ build.

After each fix lands, one branch of the temporary gate condition can
be removed.

Recommendation

Merge PR #599 (universal walker, no class-name dispatch) as the
practical resolution; this PR's value is in the documented findings
and the PJ_DESTROY_TRACE=1 debug flag, both of which support the
multi-week refCount audit that follows.

Or close this PR and let PR #599 carry the work; the design-doc
content can be cherry-picked into PR #599's branch.

Test plan

• make (build + unit tests) green.
• Refcount unit tests: walker_gate_dbic_pattern.t 21/21,
  walker_gate_dbic_minimal.t 4/4.
• Moose t/000_load.t PASS.
• DBIC walker_gate_dbic_pattern.t PASS.
• Full ./jcpan -t Moose and ./jcpan -t DBIx::Class runs —
  expected to match PR fix(walker-gate): replace class-name heuristic with universal walker check #599 numbers (61/478 Moose, 4 DBIC) since
  the runtime logic is identical.

Generated with Devin

Co-Authored-By: Devin <158243242+devin-ai-integration[bot]@users.noreply.github.com>