chore(guide): ground the workspace and fix README badges [release:v0.10.3]

[tcballard]

Summary

Two small post-release items: the workspace grounds its own agents, and the README's badges and PyPI rendering are corrected.

Adds:

• Checked-in .mcp.json serving the dogfood corpus through rac-guide (rac mcp --root .) — the team setup documented in docs/mcp.md §6, applied to this repository
• Grounding guidance in rac/prompts/rac-agent-session-start.md (imported by CLAUDE.md every session): get_summary at session start, search_artifacts before implementing, cite decisions by ID, surface conflicts instead of overriding — with a CLI fallback for sessions without MCP tools
• README: banner image URLs are now absolute (the README is the PyPI long description, where relative paths do not resolve — the banner was broken on pypi.org), and Python-version + license badges join the existing CI and PyPI badges

Scope

Included

• .mcp.json, the session-start prompt artifact, README badge row and banner URLs

Excluded

• No code, tool surface, or contract changes
• No new badges with low-signal numbers (downloads deferred until they flatter)

Verification

rac validate rac/                      # 103 valid, 0 invalid (the prompt edit is a validated artifact)
rac relationships rac/ --validate      # 0 issues
rac review rac/                        # no priority 1-2 findings
python -c "import json; json.load(open('.mcp.json'))"

The PyPI long description for 0.10.3 was checked directly: it carries the relative banner paths this PR fixes. The PyPI version badge markup was verified correct — it is dynamic (shields.io reads PyPI live, which already serves 0.10.3); any stale display is shields/GitHub-camo caching that clears on its own.

Notes For Reviewer

• The next release's long description will render the banner; the already-published 0.10.3 page keeps the broken image (PyPI descriptions are immutable per release).
• Remote agent sessions load MCP servers at session start, so .mcp.json benefits the next session, not a running one; the prompt's CLI fallback covers that case.

Implementation Process

Implemented with AI assistance under the roadmap contract. Final scope, review, and acceptance decisions were made by the maintainer.