sweep: add 17 community policies (happy-hop, step-up, academy, stilts x8, more)

[ob1-s]

Deep sweep of Hugging Face MicroDuck models against the current registry (18 entries).

Added 17 entries via pnpm uduck register (pinned revisions, hashed ONNX, resolved manifests), each with source-backed curation (no guessed normalizers/scales/hardware):

• joanfox happy-hop (publisher documents physical Pollen test 2026-09-01; claimed publisher hardware)
• Nupr-Haokun step-up, 2x Microduck Academy outputs (sprint, very-low-crouch-walk; judge kept as publisher evidence)
• 2x zhoumiaosen running (negative results preserved: 2.0/1.9 m/s targets not met)
• 8x HannesVonEssen stilts variants (10cm-200cm, exact-blob paths, hardware-mod-required)
• fffiloni polite-bow + moonwalk, koyogi run2 (no declared upstream license; flagged in notes, same precedent as jump.json)

Also bumps 3 hardcoded policy counts in tests (18 -> 35).

Verified: pnpm validate 35/35, pnpm test 20/20.

Deliberately excluded (resolver-rejected, documented for follow-up): swing (manifest schema v4), basketball (recurrent LSTM, model_api 2), q2p beak-throw (manifest schema v1), Teethyfish flamingo-ii (robot.model full_shell), forward-roll / low-crouch-walk (repos return 401). Multi-ONNX repos need exact /blob// handling per entry.