docs: AI users guide + lead README's AI section with the MCP endpoint

[rafael5]

Summary

m-dev-tools is fully AI-enabled (Phase 4 MCP server + Phase 5 continuous-enforcement gates all live). The org-landing README's "For AI agents" section pre-dated the MCP server and led with raw tools.json. This PR:

1. Rewrites that section to lead with the MCP endpoint (drop-in .mcp.json + the three tools).
2. Adds a dedicated AI users guide at docs/ai-discoverability/ai-users-guide.md for the full walk-through.
3. Leaves the direct-catalog path documented as the no-MCP fallback.

profile/README.md — rewritten "For AI agents" section

• Opens with the MCP install (.mcp.json snippet for Claude Code / Codex / Continue / any MCP-capable agent, pinned to @v0.1.0).
• Names the three tools (route_intent / describe / verify) with a one-line behaviour summary.
• Links to the new users guide for the walk-through.
• Direct-catalog path stays as a second-class fallback. Catalog links + the 8-step discovery handshake reference to AI-discoverability-architecture.md are surfaced.
• Mentions the four Phase-5 enforcement gates (freshness / link-check / license-reconcile / schema-version) so a reader knows the catalog is CI-enforced honest.
• Existing manifest-pointer bullets (m-stdlib API + grammar-surface) preserved.

New: docs/ai-discoverability/ai-users-guide.md

A consumer-facing companion to AI-discoverability-architecture.md (the latter is for framework extenders; this is for users). Sections:

1. What "AI-enabled" means — three artifacts + thin MCP wrapper + Phase-5 gates that keep them honest.
2. Two install paths — MCP via uvx + .mcp.json drop-in (Path A, recommended); raw catalog URLs (Path B) for clients without MCP support.
3. Example Claude Code session — user prompt → route_intent → describe → manifest fetch → final answer, with the key invariant "every claim is grounded in a fetched URL".
4. Agent-free shell smoke via examples/claude-code/smoke.sh.
5. The four typed-ID kinds (tool: / module: / cmd: / recipe:) + note that other kinds in the grammar aren't yet handled by describe.
6. What the agent WILL and WILL NOT do — including the "verify lists, doesn't execute" security stance with a reference to phase4-plan §3 B5 rationale.
7. Troubleshooting — uvx install, .mcp.json placement, server crash on @main, stale verified_on, offline mode.
8. Where to learn more — links to architecture doc, plan, guide, phase docs, and the MCP server repo.
9. Reporting issues — which issue tracker takes what kind of bug.

Frontmatter follows the docs-discoverability spec landed in PR #37 (doc_type: [GUIDE]; lifecycle: active; owner: rmrich5; connections: [function]).

Test plan

• make check-docs-prose clean
• make check-freshness clean
• make check-links clean
• make check-licenses clean
• make check-schema-compat clean
• make recipes-check clean
• make handshake clean
• make validate-catalog clean
• CI green