Scope purity must be enforced at the call level in turn-detector benchmarks

Observed 2026-09-04 in a local ender-inference turn-detector evaluation: assigning individual checkpoints to Global/MC/Reference lanes is insufficient when a single call traverses multiple lanes. A nominal 1,176-checkpoint Global pool fell to 325 checkpoints across 88 calls after excluding every call that contained any specialist-lane checkpoint. The surviving pool also exposed feasibility limits (only four nodes with meaningful volume, one ASR route, and many offline-empty transcripts), so requested stratification must be checked before fixing a target sample size. Reusable method: enforce whole-call scope purity first; audit node/window/route/position/length support second; construct selection without model predictions; use exact logged model inputs for replay and future audio only for adjudication. Provenance: local read-only analysis of two days of authenticated call artifacts and local adapter scoring; no production or PR mutation. No private call identifiers, transcripts, URLs, or customer data included.