Add pi package wrapper and tests for pi skill loading

[snorfyang]

PR description

pi is an agent runtime that consumes Agent Skills, but it discovers skills through its own package system: a package.json manifest whose pi.skills field maps to directories containing SKILL.md. This PR makes the repository double as a pi package, so pi users get first-class installation instead of hand-copying the skill directory:

• package.json maps pi.skills to skills/, enabling:

  pi install git:github.com/NVIDIA/nvidia-kaggle

  or a settings-based reference to a local checkout (updates with git pull):

  { "skills": ["/absolute/path/to/nvidia-kaggle/skills"] }

• tests/test_pi_package.py validates the wrapper and the skill against pi's skill rules: frontmatter name/description constraints, pi.skills path mapping to a discoverable */SKILL.md, integrity of ./-prefixed references in SKILL.md, bundled script compilation, evals manifest parsing, and that no credential artifacts are shipped

• tests/test_pi_smoke.py spawns the real pi CLI (opt-in via RUN_PI_SMOKE=1) to verify end-to-end skill discovery and /skill:nvidia-kaggle-skill routing

• README: adds a pi section alongside the existing Codex and Claude Code paths

No changes to the skill itself, the plugin manifests, or any workflow.

Checklist

• Tests pass (uv run pytest) — 71 passed, 14 skipped
• All commits are signed off (DCO) using git commit -s
• License header (SPDX-FileCopyrightText / SPDX-License-Identifier) is included on new non-markdown code files
• The change stays focused on one workflow where possible (see CONTRIBUTING.md)
• Credentials are read from KAGGLE_API_TOKEN (or another env var) — no .env, kaggle.json, or account artifacts committed
• (new/changed workflow) N/A — no workflow changed
• (plugin change) N/A — manifests untouched (claude plugin validate . passes)

🤖🤖🤖