A unit arm's parent enum is read off its carried owner, not re-looked-up by bare name (fixes NoSuchVariable for artifact_path(FleetConvergeYamlArtifact))

[gunbai-bot]

Fixes the interpreter NoSuchVariable that proud-dove-705 reported on #12101: artifact_path(FleetConvergeYamlArtifact) failed at run time whenever the closure also reached gunbc.auth.heal_publisher_provision.

How the chain was re-derived (DESIGN §6b)

1. What was observed. v1_interpreter eval_var fell through to NoSuchVariable because the reference carried no VariantValueBinding.
2. Where the binding is stamped. v1.compiler.infer infer_var_binding_kind stamps it, but only when lookup_variant_parent_enum returns an owner.
3. The earliest unjustified boundary. The arm's locals binding already carries the owner coproduct node. constructor_binding_authority says so: "the binding edge carries the owner; no name re-lookup". lookup_variant_parent_enum ignored that node and looked the owner up again by its bare name through lookup_type_for, which checks the closure-wide ident map first.
4. Why that goes wrong here. v2.std.artifact GeneratedArtifact (a record) and gunbc.generated_artifact GeneratedArtifact (a coproduct) share a name. heal_publisher_provision's closure reaches the record, so the lookup found the record, not a Disj, and no binding was stamped.

This is not an ambiguity to refuse. No module reads GeneratedArtifact bare in a position two declarations could answer, and the arm names exactly one owner. That owner's identity was simply thrown away. So the right outcome is correct evaluation, not a resolve-time refusal.

The repair

lookup_variant_parent_enum now reads the parent name off the carried owner, using local_coproduct_owner_from_locals, the helper infer_record_lit already uses for the same reason. The old bare-name lookup survives only as a fallback for arms whose locals carry no coproduct owner. The .dag is the authority; v1_compiler_infer.rs is the regenerated mirror, and a second --required-regen run shows no drift for it.

Evidence

test.claim.variant_owner_homonym_witness has two rows:

• A small fixture: a coproduct HomonymOwner beside a record HomonymOwner.
• The real specimen: gunbc.generated_artifact beside v2.std.artifact.

seed	fixture row	real specimen row
before (main)	FAIL no-such-variable: FirstHomonymArm	FAIL no-such-variable: FleetConvergeYamlArtifact
after	PASS	PASS

The original repro (a module importing heal_publisher_provision and calling artifact_path(FleetConvergeYamlArtifact)) now evaluates to .github/workflows/fleet-converge.yml.

Roster

This class is already rostered as gunbc.recurring_failure_mode a_bare_type_name_binds_the_wrong_declaration. I added a receipt for this second specimen rather than a new row. The fix covers only the value-position unit-arm route; the type-position trigger that row names is untouched.

Not in this PR

--required-regen still reports drift in v1_compiler_emit_rust.rs. That comes from main's #12034 (the item_is_effectful_callee inlining): the drift was present before this change was applied, and this change does not touch 05_emit_rust.dag.

🤖 Generated with Claude Code

[briansrls]

APPROVE-MERGE at exact head 1b9d2be0ef78c598e69942c5a696b7fdc98f37bb.

The repair is at the earliest unjustified boundary. A unit arm's locals binding already carries its owning coproduct node; lookup_variant_parent_enum now consumes that node and reads the owner's authored name rather than discarding the identity and re-looking it up through the closure-wide bare-name map. The old name-based route survives only as the fallback when no carried coproduct owner exists.

That directly restores infer_var_binding_kind's VariantValueBinding stamp, which is the fact the interpreter needed before eval_var; it does not introduce a second resolution rule or turn the legitimate homonym into a false ambiguity.

The witness discriminates both the minimal geometry (record and coproduct named HomonymOwner) and the real activation specimen (v2.std.artifact GeneratedArtifact beside gunbc.generated_artifact GeneratedArtifact), and the recurring-failure receipt correctly states that this closes only the value-position unit-arm route while leaving the existing type-position trigger open.

The .dag authority and regenerated v1_compiler_infer.rs agree. All five required checks pass and GitHub reports CLEAN. No condition remains unless the head moves.