MC pilots: exact request fidelity and structural type boundaries

Observed 2026-09-08 during a read-only collect_mc pilot review (provenance: local voice-runtime source snapshot and COR-4101 offline harness source; no call content or private evidence). A reset-context voice node whose live request keeps only the previous assistant utterance plus latest caller turn must not be evaluated or trained from a reconstructed full conversation prefix; those historical examples should remain review-only unless an exact request snapshot exists. Use a paired three-arm gate—current endpoint, same weights through candidate harness, then harness plus adapter—to separate serving/template regressions from adapter gains. Also, fine-tuning cannot repair identifier-type logic hidden in backend state: if MC versus DOT is inferred only from the latest utterance and the tool schema exposes digits only, a bare number after a DOT reprompt remains a structural backend ambiguity.