[Birfy]

Closes #878

What changed

agentdescent demo is the offline Path A entry, but it only evolved a single one-rule skill and then pointed people at python -m examples.skill_library_evolution — a command that fails outside a source checkout because examples/ is not in the wheel.

agentdescent demo --kind skill_library now builds a two-skill */SKILL.md tree (csv-total, date-parse), runs it through the same Path A machinery with the offline agent and stub reflector, and prints the usual show / apply --dry-run / init --kind skill_library next steps. --algorithm archive_selection works with that kind; text-only cards and evoskill are still refused. demo --json keeps the #862 contract.

The default demo smoke note, README, docs/quickstart-skill.md and docs/usage.md lead with the installed command. examples/skill_library_evolution.py stays as the source-checkout CLI (--dry-run, --agent, --install-to) and imports the shared fixture.

Verification at the tip

Tip SHA: 90389cd
Ran:

• pytest tests/test_demo.py tests/test_skill_library_example.py tests/test_skill_library_kind.py tests/test_guess_kind_skill_library.py tests/test_changelog_fragments.py::test_real_fragments_are_valid → 52 passed
• agentdescent demo --kind skill_library --json from /tmp (no repo cwd) → ok, kind=skill_library, held-out 0.000 → 1.000, two */SKILL.md written
  Not run:
• full suite — CI

[Birfy]

Verdict: PASS
Tip SHA: 90389cd
Ran: pytest tests/test_demo.py tests/test_skill_library_example.py tests/test_skill_library_kind.py tests/test_guess_kind_skill_library.py tests/test_changelog_fragments.py::test_real_fragments_are_valid at the tip → 52 passed; agentdescent demo --kind skill_library --json from /tmp (no repo cwd) → ok, kind=skill_library, held-out 0.000 → 1.000 over two committed rounds, csv-total/ and date-parse/ written. risk_gate.py from main → T3 (638 changed lines / 8 files exceed the T2 size cap of 250 / 12).
Not run: full suite (CI covers 3.9–3.13).

Correctness. The installed CLI now has an offline multi-skill path: demo --kind skill_library writes ≥2 */SKILL.md, runs them as a kind=skill_library EvolveSpec (same Path A compose as init), and keeps demo --json (ok, kind, rewards, learned, next). The default demo smoke no longer recommends python -m examples.skill_library_evolution, which is ModuleNotFoundError outside a checkout; README / docs/quickstart-skill.md / docs/usage.md lead with the wheel command. --algorithm archive_selection runs on the library fixture; text-only cards and evoskill are still refused before the work dir is created. The checkout example remains a thin wrapper around the shared fixture, so its existing tests still cover --dry-run / evolve_library.

Risk. T3 by size, not by kind: this is demo/docs/CLI fixture work, not a core-loop or installer change. Still needs Danyang's merge click under the T3 rule (no --auto, no --admin).