Supplements AI-discoverability-plan.md §3.4. §3.4 was authored when no
repo shipped the Phase-0 contract; with tier-1 + tier-2 onboarded
(6 of 9 repos as of 2026-05-10), that table is stale as a priority
guide.

This doc reframes priority around the actual remaining work and
ranks it by *downstream dependents* — not by ease of shipping.

Structure:
- §2 Current-state inventory (per-repo Phase-0 contract status)
- §3 Outstanding work catalogued from parent plan §7 + phase1-plan.md
- §4 Dependency graph (ASCII; shows what blocks what)
- §5 Priority ranking by downstream-leverage (P1-A ranks #1 with 6
  downstream items; tier-3 onboarding ranks #4 despite being
  "easy" because it doesn't unblock anything)
- §6 Recommended execution order — single-threaded sequence + which
  items to parallelize when bandwidth allows
- §6.4 explicit "easy first" trap: ships three quick PRs, leaves
  nothing unblocked
- §7 Concrete recommended next move: ship P1-A as a single coherent
  PR (cuts the longest dependency chain in front of the entire
  Phase 1 + 3 + 4 stack)
- §8 How to keep the doc in sync (read-then-act discipline)

Source files referenced (all on main):
- AI-discoverability-plan.md §3.4 — original tier-based framing
- phase0-plan.md — completed
- phase1-plan.md — to-do, Track A-D layout matches §3-5 here

Prose-only gate (make check-docs-prose) green: doc is .md under docs/.

Captures Phase 4 exit per phase4-plan.md §6 + §11. Mirrors
phase3-evidence.md shape: one section per gate, "What this proves"
roll-up, then each §11 done-criterion cited green.

Verified locally (gate outputs in the evidence doc):

* pytest profile/build/ — 51/51 green
* make catalog — byte-idempotent regen
* make validate-catalog — schema-strict OK
* make handshake — 8/8 steps
* make recipes-check — 4/4 clean
* uvx --from git+...@v0.1.0 m-dev-tools-mcp --tool route_intent
  --query "parse JSON in M" → ["module:m-stdlib#STDJSON"]
  (the parent-plan exit criterion, end-to-end from a fresh clone)

All nine §11 done-criteria cited green:

1. m-dev-tools-mcp repo public AGPL-3.0 + full Phase-0 contract (PR
   #1)
2. server.py exposes 3 TDD-covered tools (PR #2 — 30 dedicated
   cases)
3. v0.1.0 GitHub Release with 23-KB pure-Python wheel attached
   (Track D2)
4. tools.json carries the entry with repo_meta_url + mcp_tools_url +
   release_wheel_url (.github PR #27)
5. task_index.json has infra.agent_integration → tool:m-dev-tools-mcp
   (.github PR #27)
6. examples/claude-code/ has .mcp.json + README + smoke.sh +
   session.md template (PR #3)
7. This evidence file (PR — current)
8. uvx canonical round-trip green (run captured above)
9. CI green on every Phase 4 PR

Phase 5 (continuous enforcement) unblocked. PyPI publishing remains
deliberately deferred.

Two out-of-scope follow-ups noted in the evidence doc:

* validate-repo-meta.py UTF-8-decode bug on binary release-wheel
  exposes URLs (use --no-resolve until fixed)
* 3 Phase-3-era orphan schema fixtures gitignored but still on
  disk; auto-classifier blocks rm