Concept-anchored retrieval: wrap the embedding path (consumer 1)

Context

Consumer-1 leaf of epic #14472 — the GraphRAG pattern over the self-extracted ontology, the epic's headline retrieval consumer. Filed in the operator-directed full-fledging pass; live latest-open sweep 2026-07-02T19:37:34Z: no equivalent ticket.

#14422 OQ5 routed the retrieval contract (wrap vs replace the embedding path) to the epic's measurement floor: the #14474 concept-neighborhood read probe supplies the reachability/provenance data that decides it. This leaf implements the wrap shape — the conservative fork — with the probe artifact as its go/no-go gate.

The Problem

23,487 durable memories and 1,411 session rollups have no concept-mediated retrieval path (graduation-record healthcheck). Retrieval today is flat embedding top-k: a query about delta-updates finds texts that sound like delta-updates, but never walks delta-updates → IMPLEMENTED_BY → src/vdom/Helper.mjs → attached memories/decisions — the structural neighborhood where the institution's actual knowledge of the concept lives. Consumer 4 (measurement) cannot compute re-derivation rates until retrieval events over concepts exist.

The Architectural Reality

• Walk substrate: ai/services/memory-core/GraphService.mjs edge queries (the graduation's Matrix-1/B falsifier: metadata cannot walk; edges can).
• Concept resolution: alias-aware AFTER the defrag leaf (#14502) lands its alias→canonical map; before that, resolution hits the fragmentation split (4 ids for Golden Path — #14474 scan).
• Read-side annotations: the four-axis contract (#14422 OQ6, graduated) rendered per hop — authority (references shipped trustTier only) · fidelity · extractionProvenance (TAGGED_CONCEPT 1.0/0.8) · lifecycle. The #14474 probe establishes which axes are present in storage vs projection.
• Consumer surface: the Memory-Core query tools (query_raw_memories / KB ask_knowledge_base) — a consumed-MCP-surface change, so the Contract Ledger below is binding and the OpenAPI/tool-description sites count per the NL/MCP wiring discipline.

The Fix

One PR, wrap-shape:

1. Concept resolver: query → semantic top-k over CONCEPT-type nodes (+ alias-map lookup post-#14502) → resolved concept set with confidence.
2. Bounded neighborhood walk: 1–2 hops per resolved concept (IMPLEMENTED_BY / PARENT_CONCEPT siblings / EXPLAINED_BY / attached memories-messages-decisions), hop budget + node-type filters explicit in config.
3. Merged candidate set: embedding top-k ∪ walk-derived candidates, each walk candidate carrying provenance (via: concept-walk, the path, per-hop four-axis annotations) — the flat path stays untouched when the walk contributes nothing (wrap, never replace: embedding results are never displaced, only augmented).
4. Opt-in flag on the tool surface (conceptWalk: true or equivalent) — default OFF for one release beat; flipping the default is a follow-up decision fed by consumer-4 measurements.
5. Retrieval-event emission: each walk logs {query, resolvedConcepts, walkContributed, candidatesAdded} — the event stream consumer 4 (#14505-class measurement leaf) computes re-derivation rates from.

Contract Ledger Matrix

Decision Record impact

aligned-with ADR 0024 (read-only consumption). If the epic's OQ5 disposition later flips to REPLACE, that supersedes this leaf's wrap shape via a new decision — recorded on the epic, not silently.

Acceptance Criteria

• Wrap-shape resolver + walk module + merged-candidate path implemented, opt-in, default OFF
• Byte-identical default behavior spec-pinned (param absent ⇒ today's results)
• Walk candidates carry via/path/four-axis provenance; absent axes render as absent (degrade-by-omission, the #14466-lineage honesty contract)
• Retrieval events emitted per walk (the consumer-4 feed), schema documented
• Hermetic specs: resolver, walk budgets, merge; MCP compliance fixtures updated (both tool sites)
• Go/no-go gate honored: PR cites the #14474 artifact's reachability section; if reachability is too thin to add value, the leaf re-scopes to resolver-only and says so
• PR Resolves this leaf, Refs #14472

Out of Scope

• Replacing the embedding path (OQ5's other fork — epic-level decision).
• Anchoring writes (OQ1 leaf #14506-class); default-ON flip (measurement-fed follow-up).
• Cross-store federation (KB + MC walk merging beyond per-store enrichment).

Avoided Traps

• Replace-by-stealth — wrap means the flat path survives byte-identical under the default; the spec pins it.
• Unbounded walks — hop/type budgets are config, stated, and small; a graph walk must never dominate query latency (walk time logged per event).
• Fabricated provenance — absent axis values render absent; the probe's absent-in-storage vs absent-in-projection distinction carries into the response shape.

Sequencing

Gated on the #14474 measurements artifact (go/no-go + axis-presence reality). Alias-aware resolution upgrades automatically when #14502 lands (soft dependency; resolver ships with normalize-only fallback).

Related: #14472 (parent epic), #14474 (gate), #14502 (alias map), Discussion #14422 (OQ5/OQ6)

Origin Session ID: 8cf234b7-e698-47ca-99e2-bf865196b6aa

Retrieval Hint: concept anchored retrieval wrap embedding neighborhood walk four-axis provenance opt-in