You are reviewing a proposed RememberStack design BEFORE any runtime implementation. The owner explicitly requires your Claude Opus approval before implementation, prioritizes reliability over speed, and wants general-purpose changes rather than LoCoMo-specific score hacks.

Work read-only in /Users/jpuc/code/moje/remember-stack-retrieval-analysis-20261112. Do not edit files, run model/provider calls, probe live hosts/databases, open private environment files, or create agents. Read AGENTS.md/CLAUDE.md and relevant design instructions. Inspect the actual source; do not rely just on the proposal's characterization.

Review target (commit 431900b1d):
- plan/proposals/shared_retrieval_eligibility.md (the concrete design)
- plan/analysis/retrieval_filtering_reliability_20261112.md and readable measurements.json (evidence)
- D142/D149 and relevant current designs
- src/rememberstack/surfaces/query_engine.py
- src/rememberstack/adapters/postgres_p1.py and p1_ranked.py
- relevant entity/fact authority migrations, query-space definitions and tests

The design keeps current scoped exact ranking and broad index continuation. It fixes repeated coverage and candidate provenance, and factors fact source membership/presentation from endpoint visibility so bounded native reads can share endpoint eligibility. No runtime code is edited yet. The active benchmark's incomplete ingestion is separately preserved; do not change it.

Please check carefully:
1. Complete preservation of deletion/version/coordinate visibility, old-version-only entity provenance, document-entity bridges, merges/latest decisions, current/history identity/belief semantics, D140 evidence windows, D54 counts and contradiction-only facts.
2. Whether fact factoring is precisely specified and implementable without a second fact-truth query, permission expansion, duplicated count authority or forced full-project materialization. Recommend a smaller equivalent design if it is materially safer; identify any concrete contract gaps that would block implementation.
3. Physical work and memory at millions of rows; candidate/provenance boundaries, repeated entity work, and whether SQL materialization really ensures the stated reuse.
4. SQL snapshot/deadline/cancellation, ACL and query-space manifest/saved-query integration.
5. No paid ingestion/reuse-key changes (D56), hidden fallback, legacy runtime path, new reranker or benchmark-specific logic.
6. Whether equality, actual-plan, concurrency and full-native-operation gates are adequate and feasible. Distinguish strict correctness requirements from nice-to-have throughput benchmarks.

Return a clear APPROVED or CHANGES REQUIRED verdict. If changes are required, number blocking findings with precise document/source locations and concrete required amendments; separate nonblocking suggestions. Do not approve with unresolved blocking caveats. If you approve, briefly explain why the proposed implementation boundary is safe. Do not implement anything.


THIRD REVIEW, narrowed to your remaining four blockers from round 2. Prior reviews are committed in plan/analysis/retrieval_eligibility_reviews_20261112/. No runtime implementation has started. Re-read the revised proposal at 431900b1d and verify these amendments against source:
1. §4.3 pins the actual p9_46_0067 D140-widened evidence predicate: relation/observation evidence joins claims_visible_history UNION ALL v_memory_claim_carried_periodised. Literal new migration, existing widened fact_claim_live retained. §6 fixes hashed dependency registration. Gate1 uses a REAL database migrated to old head; no hand-picked old constants masquerade as live authority.
2. §4.2 now specifies entity-driven scope mode (reverse members -> indexed decisions -> latest across all mention decisions -> mention PK -> correlated current-content coordinate -> grouped coverage) for searches, nomination, _entity_tiers and its same-relation budget. Large unrelated-current-mention test added.
3. O(P) partition tradeoff remains explicit, no guessed time pruning. D23/postgres_schema_design §12 is reconciled at adoption; 120 monthly partitions + default required. UUID minting/reuse keys unchanged.
4. §6/gate6 name catalog_contract EXPECTED_INDEXES/EXPECTED_VIEWS, storage_capture_guard/catalog current head pins and D152 allocation/guard qualification. No old-head fallback.
Nonblocking suggestions also incorporated: offline migration refusal, pg_index validity verification, literal public/column comments, sequencing remains delivery-plan content.

Return APPROVED only if the amended design is safe to adopt and implement. If any blocking contradiction remains, name it precisely with required amendment. Do not reopen accepted architecture/tradeoffs without new source evidence; do flag real newly discovered correctness gaps. No edits, agents, model/provider/database calls or implementation. The owner is waiting on this review gate, not asking for runtime changes before approval.
