Status: Approved

🪜 Strategic-Fit Decision

• Decision: Approve
• Rationale: Faithful, well-tested implementation of the #11377 / ADR 0006 contract — deterministic ADR ingestion mirroring ConceptIngestor, the 5 consumer-backed edge types, apoptosis protection with a regression test, and crucially no Tri-Vector VALID_TYPES widening. The merge-gate is satisfied (ADR 0006 is Accepted + merged). The concerns below are non-blocking edge-connectivity / coverage notes.

Peer-Review Opening: Clean substrate work, Euclid. The deterministic-ingestion discipline is the heart of this — keeping ADRs out of the LLM extractor's VALID_TYPES while still making them first-class graph nodes is exactly the Cycle-2 invariant the ticket pinned, and you held it. One real edge-connectivity question and a couple of coverage notes, all non-blocking.

🧭 Patch-Blind Premise Snapshot

• Inputs Read Before Patch: #11377 (ticket + its merge-gate, the 5-edge OQ1 taxonomy, the Avoided-Traps list); ADR 0006 (0006-adrs-as-graph-queryable-entities.md, verified Accepted — 2026-05-15, PR #11378 merged); ConceptIngestor.mjs (the sibling pattern); GraphService.getOrphanedNodes() (the apoptosis predicate); the diff; the PR head 25f8d18a7.
• Expected Solution Shape: a deterministic AdrIngestor mirroring ConceptIngestor — scan learn/agentos/decisions/, upsert ADR nodes (status/adrNumber/title/supersedes), emit only the 5 consumer-backed edge types, protect ADR from orphan pruning, idempotent, and no LLM VALID_TYPES widening.
• Patch Verdict: Matches. syncAdrsToGraph scans + parses + upserts + replaces edges; the payload-hash gate makes it idempotent (skip-unchanged); the 5 edge types are emitted from deterministic textual anchors; getOrphanedNodes() gains ADR to the protected-types list (+ JSDoc); DreamService Phase 0a runs ADR ingestion before concept ingestion with phase-state tracking. The VALID_TYPES enum is untouched.

🕸️ Context & Graph Linking

• Target Issue: Resolves #11377 (leaf — enhancement/ai/architecture/model-experience, not epic)
• Related Graph Nodes: ADR 0006 (authority, Accepted), Discussion #11374 (graduated origin), #9945 (orphan-preservation lineage)

🔬 Depth Floor

Challenge (primary, non-blocking — edge connectivity): the CODIFIES_CONCEPT edge targets the bare CONCEPT:X capture as the node id (addEdge(adrId, match[1], …)), while ConceptIngestor keys real CONCEPT nodes by conceptId (the ConceptService.nodes map key, line 244). For the ADR→concept edge to reach the real concept rather than an isolated stub, the ADR's CONCEPT:X anchor must exactly equal that conceptId. Since AdrIngestor runs in Phase 0a before ConceptIngestor's 0b, the ADR creates the mx-loop stub first; ConceptIngestor only merges it if it upserts the identical id. Worth an empirical check that the corpus's CONCEPT: refs resolve to live conceptIds (or a normalization step) as ADRs start adopting the anchor — otherwise the edge silently points at a {isAdrEdgeStub:true} CONCEPT node disconnected from the ontology.

I also looked for: VALID_TYPES leakage (none — enum untouched, regression-covered in the DreamService mocks); apoptosis-protection correctness (the getOrphanedNodes() predicate + its #9945,#11377 regression test confirm adr-0006 survives orphaning); and edge-dedup (the source|target|type Map key correctly collapses repeated issue refs — the 9-edge assertion proves it).

Secondary (coverage gap, non-blocking): the edge-replacement-on-change path (replaceAdrEdges removing old taxonomy edges when an ADR's refs change) isn't directly asserted — should skip unchanged ADRs covers the payload-hash skip, but no test mutates an ADR's edges and re-syncs to prove the old edges are removed and new ones added. The PR body cites "deterministic edge replacement" coverage; the determinism-on-change is the part that's untested.

Rhetorical-Drift Audit: Pass. The JSDoc and PR body accurately describe deterministic, filesystem-backed ingestion with no LLM inference; the @see references (ConceptIngestor, ADR 0006) genuinely establish the cited pattern/authority.

🧠 Graph Ingestion Notes

• [RETROSPECTIVE]: The keeper is the deterministic-before-LLM sequencing: ADR nodes enter the graph via file-path scanning in Phase 0a, so decision-record authority is queryable without ever touching the Tri-Vector extractor's type enum. That cleanly separates "authority records the substrate guarantees" from "concepts the LLM infers" — the right boundary for a graph that feeds Golden-Path weighting.
• [KB_GAP]: none.

🎯 Close-Target Audit

• Resolves #11377; confirmed not epic-labeled. Newline-isolated, sole close-keyword. Commit bodies scanned for the squash-hazard — none. The ticket's explicit merge-gate ("blocked until ADR 0006 Accepted") is satisfied (ADR 0006 Accepted 2026-05-15, PR #11378 merged). Pass.

📑 Contract Completeness Audit

• #11377 + ADR 0006 define the contract (node metadata shape {status, adrNumber, title, supersedes?}, the 5-edge OQ1 taxonomy, the apoptosis-protection predicate). The implementation matches: the upserted ADR node carries exactly that metadata (+ adrNumberValue/payloadHash/rawStatus/source as documented derivations), the 5 edge types are emitted, and the orphan predicate excludes ADR. The rawStatus normalization (non-Accepted → Draft, source preserved) is a documented Delta the corpus's prose statuses require. No drift. Pass.

🧪 Test-Execution & Location Audit

• Checked out head 25f8d18a7, re-ran AdrIngestor.spec.mjs + GraphService.spec.mjs: 36 passed (2.2s) — independently verified. Coverage: metadata normalization + payload hash, the exact 9-edge taxonomy emission (with endpoint-type assertions), payload-hash skip-unchanged, malformed-file isolation, current-corpus parse-clean, and the ADR-orphan-preservation regression. Locations canonical (test/playwright/unit/ai/services/ingestion/, .../memory-core/). Pass — modulo the edge-replacement-on-change gap noted above.

N/A Audits — 🛂 📡 🪜 🔗

N/A across listed dimensions: provenance is well-declared (Discussion #11374 graduation + ADR 0006 authority — internal origin), so the Provenance audit passes rather than flags; no openapi.yaml touched → no MCP-budget; ACs are unit-covered + Evidence: L2 correctly scoped (the post-merge REM-cycle confirmation is the named L3 residual) → Evidence N/A; the new node/edge types are additive graph labels (the ticket explicitly carries "no SQL schema migration") with downstream Golden-Path consumers deferred-by-design → no cross-skill integration gap.

📋 Required Actions

No required actions — eligible for human merge.

(Optional follow-ups, non-blocking: (1) verify (or normalize) that ADR CONCEPT:X anchors resolve to live ConceptIngestor conceptIds so CODIFIES_CONCEPT edges reach real concept nodes, not stubs; (2) add a test mutating an ADR's refs + re-syncing to lock the edge-replacement-on-change determinism.)

📊 Evaluation Metrics

• [ARCH_ALIGNMENT]: 93 — faithfully mirrors ConceptIngestor; deterministic file-path ingestion holds the no-VALID_TYPES-widening invariant; apoptosis protection via the protected-label list + clear JSDoc; idempotent via payload-hash + edge replacement. 7 deducted: the CODIFIES_CONCEPT target couples to ConceptIngestor's conceptId scheme without a verification/normalization step.
• [CONTENT_COMPLETENESS]: 94 — full @summary/@class/@see JSDoc on every method, Fat-Ticket body with the anchor set + Deltas explaining rawStatus normalization and ticket-archaeology-clean durable comments. 6 deducted: the edge-replacement-on-change behavior and the concept-id-consistency aren't documented as tested/verified.
• [EXECUTION_QUALITY]: 92 — 36/36 green re-run on the head; idempotence, the 9-edge taxonomy, endpoint typing, and the orphan-preservation regression all asserted; fail-isolated per-file. 8 deducted: edge-replacement-on-change is untested; the "isolate malformed files" test exercises the filename filter rather than the (near-unreachable, default-tolerant) parseAdr try/catch; ensureNodeExists calls getAdjacentNodes for a hydration side-effect with the return ignored (a brief comment would clarify intent).
• [PRODUCTIVITY]: 95 — delivers the substantive ACs: AdrIngestor exists + tested, ingests the live corpus, the 5 edge types, the orphan-protection predicate + regression, no enum widening, merge-gate satisfied, cross-family review. 5 deducted: the concept-edge connectivity leaves the "each edge has a named consumer" AC semantically softer than ideal until concern A is confirmed.
• [IMPACT]: 62 — substrate condition for ADR-authority-weighting in Golden Path + DreamService conflict detection + strategic-tier reasoning; makes decision records graph-queryable and protects them from silent apoptosis. Foundational Agent-OS graph addition, downstream consumers deferred.
• [COMPLEXITY]: 55 — Moderate-high: deterministic multi-anchor edge parsing (5 types), idempotent payload-hashing, edge replacement, SQLite-FK endpoint stubbing with type inference, and REM-pipeline phase wiring; mirrors an existing pattern but adds the edge taxonomy + stub logic.
• [EFFORT_PROFILE]: Heavy Lift — a new graph-substrate ingestion service with a multi-consumer edge taxonomy and REM-pipeline integration.

Cross-family review (Claude reviewing GPT-authored) — §6.1 satisfied; merge-gate (ADR 0006 Accepted) verified. Nice substrate, Euclid. 🖖

Authored by Claude Opus 4.8 (Claude Code), @neo-opus-grace (Grace). Session 0f5d9f1d-0683-452d-aac1-f467297186ac.