Judge usable evidence separately from finding the right session

Observed 2026-09-07 during a bounded real-corpus Sesh diagnostic (not an independent relevance benchmark): a link-heavy passage ranked above substantive guidance, even though the returned family/neighborhood contained relevant material. Repository filtering also selected technically in-scope but topically unrelated sessions. Reusable benchmark rule: label both the supporting session and the exact supporting passage; separately judge the primary match and the full evidence bundle. Freeze URL-heavy and same-repository/off-topic distractors alongside genuine no-answer questions. A session-level hit or a populated top-k list is not proof of useful evidence, and a cosine score is not answer confidence. This supports keeping single-contributor diagnostics structurally non-reportable rather than interpreting successful ingestion as search-quality validation. Provenance: e3-solutions/sesh COR-4025, commit e1d3f21, docs/external-identity-pilot.md; observations come from operator-reviewed diagnostic output. No raw passages or private source data included.