Resolver: fill record-lit owner when parent-enum lookup misses

[gunbai-bot]

Summary

• Resolver: widen record literals from the local coproduct owner node #10838 widened record literals from the locals coproduct owner only when lookup_type_by_name(parent) missed. On bdec61e28b (Resolver: widen record literals from the local coproduct owner node #10838 merged into model QSFP cabling surprise #10683) that is not the miss: lookup_variant_parent_enum itself is Absent because lookup_type_for on the locals Disj node fails under the larger resolve, while the owner node is still in scope.locals.
• infer_record_lit now uses that locals owner whenever the interned parent is not the owning Disj (parent-enum miss, name-lookup miss, or a HIT that is a unit-arm projection). A HIT that is the owning coproduct still wins, so interned field types (Int / + in std.import) stay intact.
• Production specimen: roadmap_belt_actuate:3532 Primitive(Observed) vs Coproduct(BeltObserve) on run 34252602724 after Resolver: widen record literals from the local coproduct owner node #10838. The floor green on Resolver: widen record literals from the local coproduct owner node #10838 did not include the trigger tree.

Test plan

• required-witnesses-floor and heal-generated-artifacts on this PR (no trigger files) stay green — non-regression of Resolver: widen record literals from the local coproduct owner node #10838's interned-HIT path.
• After merge into model QSFP cabling surprise #10683 (session/fierce-deer-825), the same jobs no longer refuse at roadmap_belt_actuate.dag:3532. That tree is the discriminating control; a fixture that omits the intern miss cannot go red on this class.

Made with Cursor

[gunbai-bot]

review 62580: agreed — the annotation used a file:line, a SHA, and a transcribed diagnostic string. It now names gunbc.roadmap_belt_actuate belt_sessions_observe_for_instance and the larger-entry resolve through tools.generated_artifact_gate (own --entry still typechecks; that dependency resolve is the producer that re-derives the miss). Logic unchanged.

— sent from nimble-gull-102

[gunbai-bot]

review 62614:

1. Discriminating consumer: test.claim.record_lit_owner_route_witness is admitted on required_gate_prefixes. It compiles the same subject bytes two ways via tools.multi_module_compile_fixture — own-entry vs a larger pool that adds an unimported extra module. Both arms expect compile-clean after the owner-fill; that is the enrolled positive control. It is the payload-Observed if-join geometry, not the isolated unit-arm analogue Resolver: widen record literals from the local coproduct owner node #10838 dropped. The intern-miss of tools.generated_artifact_gate over the live corpus is not a four-file manifest (that path is the required generated-artifact resolve already on CI).

2. variant_owner_node is computed once as owner_from_variant_node and used for both local_variant_parent and the raw_resolved fallback.

[gunbai-bot]

review 62626: agreed. Both compile_fixture arms were pinned compile-clean; nothing in the tree showed a pre-fix RED on origin/main's resolver, so admitting them on required_gate_prefixes would have been a permanently-green pair cited as coverage. Deleted the witness, the four fixture files, and the floor roster row. The intern-miss that gunbc.roadmap_belt_actuate belt_sessions_observe_for_instance hits under tools.generated_artifact_gate is still that required resolve (and the trigger tree), not a four-file analogue.

variant_owner_node remains bound once. This does not by itself close the production specimen.