Turn benchmark truncation can silently discard known failure calls

Observed 2026-09-04 in local ender-inference/COR-4018 work: a broad 492-clip benchmark omitted a known Qwen failure call even though the all-call census and issue scanner contained it. The selection logic appended coverage controls, globally sorted, then truncated to the target size, so “include all detected issue checkpoints” was not actually guaranteed. A targeted Reference sequence from the omitted call also showed that no scalar threshold can separate all partial identifier turns from the final turn for the tested adapters: the MC candidate caught the final turn but prematurely finished two partial turns, while the current Reference adapter avoided early cuts but missed the final completion. Reusable method: reserve all known incidents and all detected issue checkpoints before sampling controls; evaluate fragmented identifiers as ordered sequences with partial/final labels, not independent randomly sampled utterances. Provenance: local read-only call census, lifecycle events, and hash-verified local adapter inference; no call identifiers or transcript content included.