propose: neighbors OVERRIDDEN_BY.* dot-key traversal (#165)

[HumanBean17]

Summary

• Adds in-flight propose propose/OVERRIDDEN-BY-DOT-KEY-TRAVERSAL-PROPOSE.md for #165.
• Follow-up to landed DECLARES.* neighbors dot-keys (PR feat(neighbors): navigate DECLARES.* composed edge types in one call #171 / NodeRef lacks edge-presence signals — agent cannot see client/producer/route declarations without describe #162): make override-axis edge_summary keys (OVERRIDDEN_BY, OVERRIDDEN_BY.DECLARES_CLIENT, OVERRIDDEN_BY.DECLARES_PRODUCER, OVERRIDDEN_BY.EXPOSES) navigable in one neighbors call.
• Locks design: method Symbol origin, direction="out" only, traversal via stored [:OVERRIDES] + terminal rels (no duplicate signature Cypher), via_id = overrider method on composed keys, no re-index / no ontology bump.

Test plan

• Propose-only PR — no code changes
• Reviewer approves locked decisions before implementation PR

Made with Cursor

[HumanBean17]

Critical review — propose-only PR

Verdict: Approve the propose with a few clarifications before the implementation PR. The core pivot (use stored [:OVERRIDES] instead of re-embedding IMPLEMENTS|EXTENDS + signature in the read path) is sound and already backed by existing tests (test_overrides_stored_neighbors_in_matches_override_axis_impl_ids).

---

What works well

1. Same discoverability fix as #162/#171 — edge_summary advertises OVERRIDDEN_BY.* but neighbors rejects them today; the “what you see is what you can request” principle is consistent.

2. Stored-edge traversal is the right call — Issue #165 originally framed this as embedding dispatch Cypher in neighbors. This propose correctly reverses that: override_axis_rollup_for keeps owning counts; traversal follows materialized [:OVERRIDES]. Virtual direction="out" on OVERRIDDEN_BY ↔ stored direction="in" on OVERRIDES from the declaration method is awkward but documented (locked decision #4) and matches current hint templates.

3. Parallels landed DECLARES.* — registry in kuzu_queries.py, via_id on composed keys only, edge-row counts, edge_type echoes dot-key, no re-index / no ontology bump.

4. Scope discipline — explicit out-of-scope, concrete test names, accurate supersedes list.

---

Please clarify or patch before implementation

1. Issue #165 body is stale (high)
The issue still says navigability requires signature-matching in the neighbors path. This propose’s point is that ontology 13+ [:OVERRIDES] makes that unnecessary. Please link this propose from #165 (or update the issue) when implementing so reviewers don’t re-litigate the old framing.

2. neighbors_v2 origin gate must split by axis (high)
Today all composed keys require a type Symbol (mcp_v2.py ~1738–1757). Implementation must branch DECLARES.* (type) vs OVERRIDDEN_BY.* (method). The propose says “partition by prefix” but not fail-fast vs partial success when mixing families on one origin (e.g. type + both DECLARES.* and OVERRIDDEN_BY.*). Recommend locking: reject the whole request if any requested composed key fails the origin gate, plus an explicit mixed-family test.

3. Rollup ↔ traversal equivalence must be mandatory (high)
edge_summary counts come from dispatch Cypher in override_axis_rollup_for, not from [:OVERRIDES]. The equivalence guard in the propose should be blocking for merge, not “keep or add”. Suggest: bank-chat requestAssignment + override_axis_rollup_smoke (diamond / middle-override if useful).

4. edge_filter interaction (medium)
#182 made edge_filter CALLS-only; composed keys are already rejected. Please add one line: no edge_filter on OVERRIDDEN_BY.* (same as DECLARES.*); terminal filtering is via NodeFilter only.

5. Base OVERRIDDEN_BY edge attrs (low)
Composed rows project terminal e.*; base key returns overrider methods via OVERRIDES but attrs semantics aren’t specified — lock whether attrs mirror flat OVERRIDES neighbors or stay minimal.

6. Hint test plan — success path (medium)
Grep sentinel for then neighbors(overrider_ids is good. Also add a positive hint test (mirror #171), e.g. expect neighbors(..., 'out', ['OVERRIDDEN_BY.DECLARES_CLIENT']) from describe-time templates. Note: test_hints_hv20_no_dotkey_edge_labels_in_rendered_neighbors_hints only applies to empty structural neighbors hints — not a blocker for success-path OVERRIDDEN_BY emissions.

7. Nits

• “Ontology 13+” → repo is on 14 (cosmetic).
• test_neighbors_overridden_by_dot_key_static_method_rejected — “if applicable”: either require a fixture or drop the name.

---

Merge recommendation

PR	Action
#186	Approve & merge after optional propose tweaks above
Implementation	Single PR per sequencing; move to propose/completed/ on land

Strong propose overall — main implementation risks are splitting the composed origin gate and treating rollup ↔ traversal parity as a hard invariant.