PR Review Summary

Status: Approve

🪜 Strategic-Fit Decision

Per §9 Strategic-Fit Step-Back:

• Decision: Approve
• Rationale: Correctly fixes the real residual #10058's intake surfaced — RESOLVES (PR→issue provenance) was decaying + pruning because it wasn't in the structural shield. The fix adds it to the protected set and hardens the prune step to also exclude protected types (it previously relied on them merely not decaying). Tested on its branch. The one observation (the protected set has no documented criterion) is a non-blocking nit.

Peer-Review Opening: Tight, correct provenance-integrity fix. The defense-in-depth on the prune step (not just decay) is the right call — it closes the path where a structural edge that ever acquired a low weight could still be deleted. One nit on documenting the set's criterion.

🧭 Patch-Blind Premise Snapshot

• Inputs Read Before Patch: #12644 (protect PR RESOLVES issue from decay + IssueIngestor coverage), #10058 (the stale RESOLVED_BY intake that surfaced it), the current topologyDecay decay+prune SQL on dev, and the existing structural-edge shield (IMPLEMENTS/EXTENDS/SYSTEM_TENET).
• Expected Solution Shape: Add RESOLVES to the decay-exempt structural/provenance set, and ensure the prune step also spares it (decay-exemption alone is insufficient if prune deletes by weight regardless of type). Reuse the existing decay/prune path; cover the ingestion side (ingestPullRequestFeedback → the RESOLVES edge) with a test. Must NOT protect decaying-evidence edge types (e.g. the Hebbian VALIDATES from #12640/#12643).
• Patch Verdict: Matches/improves. PROTECTED_EDGE_TYPES adds RESOLVES; both the decay UPDATE and the prune DELETE now WHERE type NOT IN (protected) via parameterized placeholders. The prune-shield is a genuine hardening beyond the ticket's literal ask. VALIDATES is correctly excluded (it's weighted-evidence, designed to decay/strengthen).

🕸️ Context & Graph Linking

• Target Epic / Issue ID: Resolves #12644
• Related Graph Nodes: #10058 (originating intake); GraphService.topologyDecay; IssueIngestor.ingestPullRequestFeedback; the RESOLVES edge; permanent-provenance vs decaying-evidence (VALIDATES #12640/#12643) edge classes.

🔬 Depth Floor

Challenge (non-blocking nit): PROTECTED_EDGE_TYPES now encodes a real conceptual class — permanent structural/provenance edges (IMPLEMENTS/EXTENDS/SYSTEM_TENET/RESOLVES) — but the list carries no documented criterion. The live discriminator matters right now: the freshly-added VALIDATES edge (#12640/#12643) is deliberately not protected, and that's correct — it's Hebbian weighted-evidence (repeated cycles strengthen it; weakening evidence should decay), not permanent provenance. But a future edge-type author can't infer that rule from the bare list. Suggest a one-line comment on the constant naming the criterion ("permanent structural/provenance — never decays/prunes; contrast weighted-evidence edges like VALIDATES which decay by design"), so the next addition is classified deliberately rather than by guess. I also checked the SQL: parameterized placeholders (no injection), and the prune MAX(...,0.1) floor interplay is consistent.

Rhetorical-Drift Audit (per guide §7.4):

• PR description: framing ("decay + pruning protection") matches the two SQL guards.
• Anchor & Echo: the comment update ("durable structural/provenance edges completely from decay and pruning") accurately describes the change.
• [RETROSPECTIVE]: none.
• Linked anchors: #10058 is cited as the originating intake (non-close-target), accurate.

Findings: Pass.

🧠 Graph Ingestion Notes

• [KB_GAP]: N/A. (Noted: the PR body says the KB semantic ticket-sweep failed — collection unavailable — and the author fell back to live GitHub + Memory Core + source. Correct fallback, not a gap in this PR.)
• [TOOLING_GAP]: the KB collection being unavailable for the pre-filing sweep is a standing tooling-availability gap (not introduced here).
• [RETROSPECTIVE]: Provenance edges (RESOLVES) joining the permanent-shield class is the right model — a PR did resolve an issue, permanently; that fact shouldn't fade with ambient Hebbian decay the way evidence-strength edges do.

🎯 Close-Target Audit

• Close-targets identified: #12644
• #12644 labels: enhancement, ai, testing, architecture, model-experience — NOT epic. Valid leaf; Resolves #12644 correct keyword.

Findings: Pass.

N/A Audits — 📑 🪜 📡 🔗

N/A across listed dimensions: internal graph-decay logic (no public/consumed contract surface — 📑), covered by unit specs (🪜 Evidence), no openapi.yaml touch (📡), no new cross-skill convention (🔗).

🧪 Test-Execution & Location Audit

• Branch checked out locally (codex/12644-resolves-edge-hardening); verified the diff is present (PROTECTED_EDGE_TYPES ×4, RESOLVES in the decay/prune guards).
• Ran GraphService.spec.mjs + IssueIngestor.spec.mjs --workers=1: 38/38 pass (incl. the new decay/prune-protection + the IssueIngestor PR→RESOLVES-edge ingestion coverage).
• Location: new assertions live in the existing canonical specs.

Findings: Tests pass.

📋 Required Actions

No required actions — eligible for human merge.

(Non-blocking nit in the Depth Floor: document the PROTECTED_EDGE_TYPES criterion so future edge-types — esp. weighted-evidence ones like VALIDATES — are classified deliberately.)

📊 Evaluation Metrics

• [ARCH_ALIGNMENT]: 92 - RESOLVES correctly joins the permanent-provenance shield; prune-step hardening is proper defense-in-depth; parameterized SQL. 8 deducted: the protected-set criterion is undocumented (the permanent-provenance vs decaying-evidence distinction is load-bearing now that VALIDATES exists).
• [CONTENT_COMPLETENESS]: 90 - Decay + prune + ingestion all covered by tests; comment updated. 10 deducted: no criterion comment on the constant (why RESOLVES in, VALIDATES out).
• [EXECUTION_QUALITY]: 93 - 38/38 on the actual branch; correct parameterized placeholders; the prune-shield closes a real gap. 7 deducted: minor — relies on the reader inferring the protected-class semantics.
• [PRODUCTIVITY]: 95 - Delivers #12644 (decay protection + ingestion coverage) + the residual #10058 surfaced; 5 deducted for the undocumented criterion.
• [IMPACT]: 55 - Protects PR→issue provenance integrity that close-target / lifecycle graph queries depend on; contained but real.
• [COMPLEXITY]: 38 - Moderate-low: a frozen protected-set constant + two parameterized SQL guards + ingestion-side test scaffolding (sync-fs mocks).
• [EFFORT_PROFILE]: Quick Win - Targeted provenance-integrity fix reusing the existing decay/prune path; high integrity-ROI at low complexity.

Correct provenance-integrity fix with proper defense-in-depth — approve. Just give the protected set a one-line criterion comment so the next edge-type (especially a decaying-evidence one) lands on the right side of the line. 🖖

— @neo-opus-grace (Claude Code)