Add CLI characterization test suite for AI flags

[leochong]

This adds a generate-in-CI characterization test that pins observable CLI output of migra.command.run() for every AI-powered flag combination, so future refactors get caught by a diff against a freshly-generated baseline — without committing any baseline files to the repo.

Files added:

• tests/characterization/scenarios.py — 21 scenarios covering
  error-path + empty-diff + rich-output for --explain, --rollback,
  --advise, --generate, --explain-drift, --output json, and combined
  flags
• tests/characterization/capture.py — capture engine with
  run_scenario() / capture_all()
• tests/characterization/test_capture.py — 7 pytest tests (all pass
  without Postgres for non-fixture scenarios)
• tests/characterization/README.md — documentation
• scripts/compare_characterization.py — cross-run drift comparison
• scripts/download_base_artifact.py — GitHub REST API artifact
  downloader for cross-branch comparison
• .github/workflows/ci.yml — new 'characterize' job
• .gitignore — exclude _artifacts/

Key design:

• No live Anthropic/Postgres dependency beyond existing mocks
• Non-blocking: continue-on-error: true, no --strict by default
• No new third-party GitHub Actions
• --from-file scenarios with real fixtures (everything/, enumdeps/) capture full migration SQL + AI output on runs with Postgres