This evaluation answers one narrow architectural question: does the serving path need semantic vectors now, or can the smallest auditable retrieval design meet its declared quality and safety floor?

The versioned fixture declares labels and thresholds separately from runtime results. A vector serving dependency is justified only when bounded lexical plus one-hop relation retrieval falls below Recall@5 0.90. Changing ranking code may produce a new versioned report, but must not rewrite the existing labels or threshold to manufacture a pass.

Fixture v1 fixes the policy at five final results, minimum lexical query-token coverage 0.40, and eight exact relation-target reads. The corpus contains:

• six direct lexical queries;
• three queries whose relevant procedure is reachable through an explicit RelatedTo edge from a lexical anchor;
• one known paraphrase with no token overlap or curated edge;
• an Active conflict target, a hidden Candidate, and an exact matching node in a foreign namespace.

The source is core/tests/fixtures/durable-memory-retrieval-v1/corpus.json.

Recall@5 is the fraction of labeled queries with a relevant node in the first five results. Mean reciprocal rank (MRR) averages the reciprocal rank of the first relevant result. Each fixture-v1 query has one independently declared relevant node.

Relation expansion recovers all three explicitly linked procedures. The remaining miss is a paraphrase, which is retained as evidence of the known lexical limitation rather than relabeled away.

The public-API integration test also proves that retrieval:

• returns only Active nodes from the exact tenant/principal/scope namespace;
• never expands ConflictsWith or recursively traverses the graph;
• respects both the final-result and exact-read bounds;
• leaves admission and use counters unchanged during preview_recall;
• still requires successful admission of the current revision before any hit enters model context.

These are serving invariants, not quality metrics. A higher recall score cannot compensate for violating one of them.

Fixture v1 meets the predeclared 0.90 relation Recall@5 gate, so A3S Code does not add embeddings, a vector index, model egress, or vector lifecycle state to this path. This synthetic result is a deterministic correctness gate, not a claim about every production distribution. Hosts should retain shadow measurement and add a new versioned, independently labeled corpus when real misses show that the decision should change.

The separate Durable Memory Product Evaluation reuses these labels through real AgentSession turns and adds write precision, evidence fidelity, conflict preservation, context, call, nominal cost, and admission gates. Neither deterministic test substitutes for host qualification on a representative production distribution.

Run the report from the Code crate workspace:

cargo test -p a3s-code-core --test durable_memory_retrieval_eval -- --nocapture