fix: replace class-name walker-gate heuristic with property-based gate

[fglock]

Summary

Removes the hard-coded class-name walker-gate heuristic
(Class::MOP/Moose/Moo) and replaces it with a property-based
gate that captures the underlying lifetime semantics directly.

The Problem

When PerlOnJava's cooperative reference count for a blessed object
transitions to 0, MortalList.flush() fires DESTROY. But cooperative
refCount can transient-zero mid-statement during deferred-decrement
processing — even when the object is genuinely held by a long-lived
owner (like our %METAS).

Master's workaround was a class-name heuristic gate: if the class is
in Class::MOP/Moose/Moo, consult the reachability walker before
firing DESTROY; otherwise, fire normally. This worked but was
unsatisfying — it baked specific class names into the runtime and
didn't extend to other module-global metadata.

The Solution

Capture the underlying property: an object is "module-global metadata"
if it's stored as a value in a package-global hash (e.g., $METAS{Foo} = $meta).

// New on RuntimeBase:
public boolean storedInPackageGlobal = false;

// New on RuntimeHash:
public boolean isGlobalPackageHash = false;  // set when registered
                                             // in GlobalVariable.globalHashes

// In RuntimeBaseProxy.set, when this is a RuntimeHashProxyEntry:
if (parent.isGlobalPackageHash && lvalue.value instanceof RuntimeBase rb) {
    rb.storedInPackageGlobal = true;
}

Walker gate at MortalList.flush and the mirror site in
RuntimeScalar.setLargeRefCounted now checks base.storedInPackageGlobal
instead of classNeedsWalkerGate(base.blessId).

For DBIC rows: never stored in a package-global hash → flag never set
→ standard refCount → DESTROY on undef $row correctly fires.

For CMOP metaclass: stored in $METAS{...} → flag set → walker
checks reachability → rescue across transient zeros.

Verification

All green WITHOUT the class-name heuristic:

Test	Status
Unit tests (make)	✅ PASS
use Class::MOP; use Moose;	✅ PASS
examples/moose.pl (full Moose demo, 7 test groups)	✅ PASS
DBIC t/52leaks.t	✅ 11/11 PASS

Performance

• Per-store: one boolean check + one flag set if destination is in
  a package-global hash. O(1); zero-cost for stores not into package
  globals.
• Walker: same gating as the old heuristic — only fires when
  object has weak refs AND storedInPackageGlobal AND refCount→0.
  Same performance envelope.
• Diagnostics: zero overhead when env-flags off (single boolean
  check per call site).

What's Included

This unified PR includes:

1. The fix (D-W6.18): property-based walker gate replacing the
   class-name heuristic.
2. RefCount discipline audit (D-W6.13): paired
   releaseActiveOwner calls at all known direct --refCount sites
   (RuntimeList, Storable, RuntimeArray.pop/shift,
   RuntimeStash, DestroyDispatch, WeakRefRegistry.weaken,
   etc.).
3. Walker improvements (D-W6.14/16):
   ReachabilityWalker.isScalarReachable walks roots looking for
   scalar identity (existing walker only matched RuntimeBase
   referents); skips weak refs; follows closure captures.
4. Per-RuntimeBase ownership tracking (D-W6.13): activeOwners
   set + reachableOwnerCount() infrastructure ready for future
   use.
5. Diagnostic env-flags (D-W6.10/11/12): PJ_REFCOUNT_TRACE,
   PJ_WEAKCLEAR_TRACE — gated; zero-cost off; used for the
   diagnostic work that led to this fix.
6. Regression tests: src/test/resources/unit/refcount/drift/
   (our_metas_underflow.t + cmop_pattern.t).
7. Documentation: dev/modules/moose_support.md D-W6.7 through
   D-W6.18 — full investigation, alternatives considered, and
   rationale for the chosen approach.

Why This Is Correct

The heuristic worked because it captured a real property —
"objects whose lifetime is module-global metadata". This PR moves
from "approximate by class name" to "exact via flag set at store
time".

• Class-agnostic: works for any module that stores metadata in
  package-global hashes (extensibility).
• Behaviorally equivalent to the old heuristic for tested
  modules.
• Captures the underlying property explicitly rather than via
  ad-hoc class-name list.

Alternatives Considered (and Rejected)

• Per-statement FREETMPS-style flush: doesn't fix the
  underlying refCount imbalance — flush schedule changes WHEN
  decrements fire, not WHETHER they're paired (D-W6.17).
• Universal walker (no class filter): regresses DBIC's leak
  detection by 4-9 tests due to phantom owners (D-W5/W6.16).
• Find unpaired refCount site: bounded but unclear if findable
  in practice; even fixed, the heuristic remains (orthogonal).
• Java-GC ground truth: major rewrite, too risky for the
  current iteration.

Follow-Up

The classNeedsWalkerGate method remains in DestroyDispatch.java
(used only by the D-W6.10 diagnostic tracer at bless time).
Production code paths no longer reference the class-name list.
A follow-up cleanup commit can delete it entirely once diagnostic
activation is reworked.

Test plan

• make passes (full unit test suite)
• use Class::MOP; use Moose; works WITHOUT heuristic
• examples/moose.pl passes (full Moose demo)
• DBIC t/52leaks.t 11/11 passes
• New regression tests pass

Generated with Devin