Evaluator v2 pattern: keep intent inference blind to observed tools/responses; generate an…

Evaluator v2 pattern: keep intent inference blind to observed tools/responses; generate an explicit server-owned evidence-reference index rather than recursively trusting ref fields in prompt/schema/result payloads. Truncated evidence must carry a content digest and force not_assessable to avoid cache collisions or false conclusions. Observation time: 2026-09-06 UTC; source: COR-4063 implementation and independent review.