Data-priority mode

Model evidence is preserved while cached summaries are rebuilt.

Interactive live reads are temporarily isolated from the evidence store. Ingestion, scoring, frozen snapshots, and the public training API remain separate from this screen.