plan: CALLS noise + resolution (propose rev5 + PLAN + prompts)

[HumanBean17]

Summary

• Updates propose/CALLS-NOISE-AND-RESOLUTION-PROPOSE.md to revision 5 (status: under review) — fixture anchors, HV37 wording, and plan alignment after #179 review.
• plans/PLAN-CALLS-NOISE.md — 3-PR landing order, pinned fixture anchors, named tests, per-PR sentinel greps.
• plans/CURSOR-PROMPTS-CALLS-NOISE.md — per-PR Cursor handoffs (PR-1 → PR-3). Do not open PR-1 code until this PR merges.

Builds on merged propose #178 (revision 3 on master). This branch adds revision 4 implementation-contract patches plus revision 5 anchor/prompt fixes.

Tracks #177.

What changed in revision 4 (propose)

• ORDER BY on neighbors(out, ['CALLS']) — source-order transcript at MCP, not just in the graph.
• edge_filter pushdown — filter in Cypher before limit, not post-fetch truncation.
• exclude_external stays on find_callers only; JDK noise on neighbors uses edge_filter.
• HINTS/ontology PR-3 checklist, supertype-dedup pseudocode, /mini-map cross-link for accessor noise.

What changed in revision 5 (this branch)

• Fixture anchors: ClientMessageProcessor#process (bank); call_graph_smoke for supertype dedup + overload_ambiguous (not fictional OrderService / MyRepository).
• HV37 awkward-case footnote: exclude_external only on find_callers/find_callees; OTHER role filter bluntness documented.
• Test names unified (pushdown_in_cypher, perf test ..._client_message_processor, heavy-gated).
• PR-2: unresolved-calls CLI deferred to PR-3 (no empty stub).
• PR-3: interleave tie-break + dedup_calls canonical line in plan.
• CURSOR-PROMPTS-CALLS-NOISE.md — iteration-loop tests, sentinels, scope per PR.

Plan (no code in this PR)

PR	What	Breaking?
PR-1	callee_declaring_role on CALLS + supertype dedup + counters	Re-index, ontology 15; row-count delta at dedup sites
PR-2	edge_filter on neighbors_v2 + ORDER BY + pushdown	No
PR-3	Move phantom/chained receiver failures to UnresolvedCallSite	Yes

Test plan

• Propose + plan + prompts review only (no runtime tests in this PR)
• After merge: implement PR-1 → PR-2 → PR-3 per plans/CURSOR-PROMPTS-CALLS-NOISE.md

[HumanBean17]

Review (revision 4 + plan)

Verdict: approve with doc fixes before merge. This is a solid implementation contract on top of merged #178. I would not start PR-1 until fixture anchors and CURSOR-PROMPTS-CALLS-NOISE.md are settled.

What looks good

• ORDER BY + pushdown (Decisions 36–37) match real code: neighbors_v2 flat CALLS path has no ORDER BY and slices in Python (mcp_v2.py ~1391–1459). Locking ORDER BY e.call_site_line, e.call_site_byte with predicates before offset/limit is the right fix for #177-style truncation.
• Phantom split line refs are accurate: remove chained_receiver (1190–1200) and unresolved-receiver phantom (1202–1212); keep known-external (1257–1271) and overload_ambiguous (1284–1289).
• §3.9.1 HINTS checklist is necessary — FUZZY_STRATEGY_SET still includes phantom / chained_receiver; PR-3 will break hints without H1–H8.
• 3-PR landing order (additive schema → MCP filter → breaking graph) is sound.

Blockers (please fix in this PR)

1. Fixture / test anchors are fictional or absent on bank-chat-system:

   • OrderService.process does not exist in the fixture (HV1, HV21, HV34, perf test name). Pin a real method — e.g. ClientMessageProcessor#process (~57 outbound CALLS on a fresh bank index).
   • test_pass3_supertype_dedup_jpa_repository_save_one_row (MyRepository extends JpaRepository) — no such type in bank; also zero same-site duplicate save CALLS today. Use tests/fixtures/call_graph_smoke/ or a small dedicated fixture per tests/README.md.
   • test_pass3_overload_ambiguous_still_n_rows on bank — overload_ambiguous count is 0 in bank graph; needs a dedicated fixture, not bank alone.

2. Internal contradiction on exclude_external: Decision 38 / §3.4.2 correctly say it is not on neighbors, but §4.5 HV37 note still says agents can use exclude_external=True for library noise without scoping to find_callers / find_callees. Please fix that sentence (and grep for similar slips).

3. Test name drift: propose §3.4.1 has test_neighbors_calls_edge_filter_pushdown_role; plan PR-2 has test_neighbors_calls_edge_filter_pushdown_in_cypher — pick one verbatim name in both files.

4. Plan header: still points at #178 as the active propose PR; #178 is already merged. Update to #179 / “rev4 in this PR”.

5. Missing CURSOR-PROMPTS-CALLS-NOISE.md: plan defers until PR-1. Repo norm is plan + prompts together (plans/completed/CURSOR-PROMPTS-*.md). Please add a skeleton prompt file in this PR or open a follow-up before any code lands.

Non-blocking (implementation / prompts)

• Post–PR-3 JDK noise via exclude_callee_declaring_roles: ['OTHER'] is blunt (known-external rows are also OTHER) — ensure AGENT-GUIDE states this clearly.
• Perf test 1.5× median will flake in CI unless gated (JAVA_CODEBASE_RAG_RUN_HEAVY or skip) — plan says “optional heavy” but also lists under deliverables; resolve in prompts.
• Prefer deferring unresolved-calls CLI stub to PR-3 (empty output before UCS tables is misleading).
• PR-3: specify pagination / tie-break for include_unresolved interleave and canonical line when dedup_calls=True.
• PR-2 sentinels: treat rg "LIMIT.*call_site" as advisory; rely on named tests for pushdown-before-limit.
• PR-1 still changes row count via supertype dedup — call that out in README re-index notes (not only “additive column”).

Merge checklist

• Replace OrderService.process with real bank method IDs (or label as illustrative + pin test IDs).
• Fixture paths for supertype dedup + overload_ambiguous tests.
• Fix §4.5 HV37 exclude_external wording.
• Unify pushdown test name; update plan header (#178 → #179).
• Add CURSOR-PROMPTS-CALLS-NOISE.md or explicit follow-up before code PRs.

Happy to re-review after a small doc pass.