docs(D-W6.6): precise Class::MOP failure stack — Try::Tiny + _post_add_attribute

[fglock]

Summary

D-W6.6 narrows the Class::MOP drift failure to the exact line in
_post_add_attribute — and shows that the visible error is a CASCADE
caught by Try::Tiny.

What the failure stack actually shows

A $SIG{__DIE__} probe before use Class::MOP (gate disabled)
captures:

FAILURE at:
  main::__ANON__         at .../Try/Tiny.pm:139
  (eval)                 at .../Try/Tiny.pm:140
  Try::Tiny::try         at .../Class/MOP/Class.pm:897
  Class::MOP::Class::_post_add_attribute
                         at .../Class/MOP/Mixin/HasAttributes.pm:41
  Class::MOP::Mixin::HasAttributes::add_attribute
                         at .../Class/MOP.pm:188

So the bootstrap chain is:

# CMOP.pm:188
HasMethods->meta->add_attribute(
    Attribute->new('wrapped_method_metaclass' =>
        reader => { 'wrapped_method_metaclass' => \&CV },
        ...
    )
);

# HasAttributes::add_attribute (line 41)
$self->_attach_attribute($attribute);   # weaken back-ref
$self->{_attribute_map}{...} = $attr;
$self->_post_add_attribute($attribute);

# Class.pm:897 — _post_add_attribute body
try {
    local $SIG{__DIE__};
    $attribute->install_accessors;     # ← dies (associated_class undef)
} catch {
    $self->remove_attribute(...);       # ← then THIS dies for the same reason
    die $_;
};

The visible error ("Can't call method 'get_method' on an undefined
value at Attribute.pm line 475") is the SECOND failure — the catch
block's call to remove_attribute → remove_accessors. The first die
is hidden by Try::Tiny's local $SIG{__DIE__}.

What landed

• src/test/resources/unit/refcount/drift/cmop_add_attr_loop.t —
  the exact shape: 11 iterations of _attach_attribute →
  _attribute_map insert → _post_add_attribute with HASH-form
  reader. Both Perl 5 and PerlOnJava (gate disabled) pass.
• try_tiny_weak.t — Try::Tiny + weakened back-ref alone
  (4 patterns). Pass.

So the simple shape isn't enough. The bug requires the recursive
bootstrap where HasMethods->meta is itself a Class::MOP::Class
with a multi-level @ISA chain (Module, HasAttributes, HasMethods,
HasOverloads) currently being built up.

Next concrete probe

A $SIG{__DIE__} Perl-side probe that prints
$self->associated_class // 'UNDEF' immediately before each die.
The first die where associated_class is 'UNDEF' reveals the
exact LAST PRIOR refCount-decrement step that took the metaclass to
0. That decrement's stack identifies the specific linearized_isa /
_method_map traversal inside install_accessors that loses the
metaclass strong hold during the partial-build state.

Test plan

• make (build + unit tests) green.
• cmop_add_attr_loop.t passes on master.
• cmop_add_attr_loop.t passes with the gate disabled.
• try_tiny_weak.t passes both ways.

Open D-W6 PR backlog (cumulative)

#	Branch	Lands
599	feature/moose-phase-d-v2	Universal walker (no class-name dispatch)
603	fix/d-w6-1-sub-install-drift	sub_install.t
605	fix/d-w6-2-closure-capture-drift	closure_capture.t + hash_slot.t
606	fix/d-w6-4-pending-double-add	weak_metaclass.t
607	fix/d-w6-pending-instrumentation	Diagnostic env-flags
609	fix/d-w6-function-hash-store	function_hash_store.t + cmop_bootstrap.t
this PR	fix/d-w6-bisect-cmop	try_tiny_weak.t + cmop_add_attr_loop.t + precise stack

Generated with Devin

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