add plan and cursor prompts for schema v2

[HumanBean17]

Summary

• Adds plans/PLAN-SCHEMA-V2.md — four-PR execution plan implementing propose/SCHEMA-V2-PROPOSE.md (EDGE_SCHEMA infra, HTTP_CALLS Client→Route flip, Producer node + ASYNC_CALLS flip, hints v3 gated on HINTS-V3-PROPOSE).
• Adds plans/CURSOR-PROMPTS-SCHEMA-V2.md — per-PR Cursor handoffs with sentinels, named tests, and manual evidence commands.

Satisfies Decision 29 merge gate: plan + prompts must land before PR-A code.

Test plan

• Docs-only change; no code or pytest impact
• Plan reviewed against propose §6 migration and 32 decisions

Made with Cursor

[HumanBean17]

Review summary

Verdict: Approve — satisfies Decision 29 (plan + prompts before PR-A code). Docs are well structured, aligned with propose/SCHEMA-V2-PROPOSE.md, and match the repo’s PLAN-* / CURSOR-PROMPTS-* pattern. CI green; scope is appropriately docs-only.

Below are items to resolve before or during PR-A implementation (not blockers for merging this PR).

---

Critical (before PR-A code)

1. Decision 30 vs plan wording

The propose says HINTS-V3-PROPOSE.md must exist before PR-A merges. The plan softens this to “tracked plan to land PR-D.” propose/HINTS-V3-PROPOSE.md does not exist yet. Merging PR-A code without HINTS-V3 scheduled leaves UC3 (neighbors(method, out, HTTP_CALLS) → [] with no actionable hint). Recommend either opening the HINTS-V3 propose in the same stack before PR-A code, or tightening the plan to match the propose.

2. ONTOLOGY_VERSION bump in PR-A breaks v13 indexes immediately

kuzu_queries.KuzuGraph.get() rejects graphs where meta.ontology_version < ONTOLOGY_VERSION. Bumping to 14 in PR-A (with unchanged graph shape) will refuse all v13 indexes as soon as PR-A lands — not only after PR-B/C as the propose/README imply (“rebuild after B+C”).

Pick one and document explicitly in PR-A README:

• A: Defer version bump to PR-B (first shape change), or
• B: Keep bump in PR-A but state “full re-index required after PR-A” (version-only break).

3. PR-A manual evidence uses bare python -c

CURSOR-PROMPTS common rules require .venv/bin/python; the PR-A manual block should use .venv/bin/python -c for consistency with repo rules.

---

Medium (during implementation)

• p: vs producer: — Decision locks p: (parallel c:). Plan PR-C mentions producer: prefix; mirror Client’s dual acceptance (c: / client:) in the plan.
• BROWNFIELD_RESOLVER_STRATEGY_SET — enumerate members explicitly or add a test that fails on unknown emitted strategies (avoid drift into PR-D).
• PR-B half-migration — async trace/pr_analysis may stay one-hop until PR-C; ensure PR-C adds explicit async pr_analysis coverage (UC9 async arm).
• Doc generator CI — pick one enforcement path (pytest subprocess --check is enough); avoid “optional workflow” with no pytest hook.
• CURSOR-PROMPTS PR-D — add test_hints_neighbors_client_edge_summary_nonzero_out (in plan, missing from prompt).
• PR body “32 decisions reviewed” — plan table doesn’t enumerate all 32; optional decision→PR mapping appendix would help code PR reviewers.

---

Nits

• PR-A must ship 10 pre-flip EDGE_SCHEMA entries — do not paste propose Appendix A (11 post-flip entries) into PR-A.
• Propose Decision 17 (“hints gated on PR-A”) wording conflicts with Decision 30; plan PR-D blocking is the operative rule.

Happy to merge this planning PR as-is; the above is for the implementation chain (A→B→C→D).