Repair Search Query, Ranking & Result Integrity Bugs

The likely issue is a contract mismatch around search query parsing, filters, facets, ranking, sort stability, permissions, pagination, relevance signals, tie-breaking, empty queries, stale candidates, duplicate results, and result explainability. I would first define the invariant: Search results must be drawn only from the authorized eligible corpus and ordered by one explicit query/ranking contract that remains stable across filters, pagination, retries, and tie cases. Then I would trace the smallest owning boundary, keep VERIFIED facts separate from LIKELY/UNKNOWN state, apply or propose the narrowest repair, and add deterministic normal, boundary, failure, retry/cancellation, and recovery tests. The workflow preserves authorization, privacy, security policy, public interfaces, user-visible behavior, and documented ownership. With repository/workspace access, findings can map to exact paths and focused local validation; with partial evidence, the result remains an explicit diagnosis/evidence map with no invented execution claims.