Separate semantic extraction from action judgment in call evaluation

Observation (2026-09-07; source: ender-inference COR-4058 / PR https://github.com/e3-solutions/ender-inference/pull/134 and follow-up requirements analysis): A scalable call evaluator should not run one holistic LLM over every full trace. First compile immutable typed evidence from universal_context, then perform blinded semantic extraction (intent inferred without seeing the chosen tool), and finally dispatch independent deterministic/JUDGE/HYBRID checks over bounded node context plus correlated downstream results. Aggregate only across compatible prompt/schema/model versions; publish quality score and evidence coverage separately. Random cohorts must persist filters, seed, and resolved call-ID manifest for reproducibility.