[Feat]: PyPI wheel smoke test — load every registry skill after pip install

[rosspeili]

Feature Description

Clone-repo tests (pytest tests/, pytest skills/) do not prove that a built PyPI wheel ships all registry bundles correctly. Packaging regressions (__init__.py, MANIFEST.in, package-data) can slip through until a user runs pip install skillware in a clean project.

Add a CI job (or step after build) that:

1. Builds the wheel/sdist from the PR branch (python -m build or pip wheel .).
2. Creates a fresh venv with no editable install.
3. pip install the built artifact with [all] (or minimal + per-skill extras as needed).
4. For each skills/**/manifest.yaml, runs SkillLoader.load_skill("<category>/<name>") and asserts bundle keys (manifest, instructions, module).
5. Optionally runs one offline assertion per skill (manifest name matches folder id).

Parent context: RFC #156 testing architecture. Complements bundle tests in editable mode; does not replace pytest skills/.

Rationale

Skillware's core promise is pip install skillware → load skills by registry ID. Wheel-only breaks are rare but painful (missing __init__.py, assets not in wheel, wrong package-data). A smoke test catches them before release and gives contributors confidence that packaging changes are safe.

This is higher leverage than adding more maintainer tests under tests/skills/ because it validates the shipping path agents and PyPI users actually use.

Implementation Idea

• Add .github/workflows/wheel-smoke.yml (or extend ci.yml with a wheel-smoke job).
• Script: scripts/wheel_smoke_test.py invoked from CI.
• Use dummy env vars already set in CI (ETHERSCAN_API_KEY, ANTHROPIC_API_KEY) where load triggers warnings only.
• Document in docs/TESTING.md under a short "Packaging smoke test" bullet.
• CHANGELOG [Unreleased] when merged.