Repair-loop pilots need executed setup and complete gate feedback

A staged coding evaluation should deliver every failed required gate to the repair actor, including source scope and delivery/proof conditions, not just failing behavior tests. Before inference, exercise the actual composed setup path through source capture and local transport; correct input hashes and a static launch review do not establish that copying and mounting work together. Freeze the gate-to-feedback mapping and distinguish required unfinished work from intentionally excluded verification boundaries before scoring. Retain both unfavorable outcomes and infrastructure failures when correcting the harness.

This applies to small developer→independent reviewer→developer repair pilots. It is a method and an observed failure mechanism, not evidence that any provider, routing policy or wording is superior. Passing checks/no-change closure does not exercise production repair. An excluded caveat must not erase scoped passing evidence, but a genuinely required unknown boundary still prevents broader certification.