feat(plugins): add built-in code-puppy-agent architecture skill

[thomwebb]

What

Adds a new built-in plugin, code_puppy_agent, that ships a SKILL.md — a self-awareness document teaching the LLM Code Puppy's internal architecture (agents, tools, plugins, models, MCP, sessions, skills) so it can navigate the codebase and extend the product without guessing. This mirrors the pattern used by Hermes's hermes-agent skill.

Why

When the LLM needs to add a plugin, debug MCP, or understand tool registration, it currently has to re-derive the architecture by reading source. A skill that documents the actual internals — hook names, discovery precedence, prompt layering — turns guessing into lookup.

What changed

File	Change
code_puppy/plugins/code_puppy_agent/__init__.py	NEW — package init
code_puppy/plugins/code_puppy_agent/register_callbacks.py	NEW — registers the skill via register_skills callback (skill_md_path)
code_puppy/plugins/code_puppy_agent/SKILL.md	NEW — 14-section architecture reference (~19 KB)
tests/plugins/conftest.py	MODIFIED — marker-based _isolate_plugin_skills fixture so the built-in skill doesn't leak into filesystem-only discovery tests
tests/plugins/test_agent_skills.py	MODIFIED — 2 plugin-skill tests opt in via @pytest.mark.plugin_skills

Design notes

• skill_md_path over inline — the SKILL.md is ~19 KB; inlining via skill_md would bloat register_callbacks.py. Reading the file via path is cleanest and preserves hand-authored frontmatter.
• Test isolation — the autouse fixture checks request.node.get_closest_marker("plugin_skills") and skips the monkeypatch when present, so plugin-skill tests declare their intent at the decorator instead of undoing a fixture yield.
• Name consistency — the entry "name" matches the SKILL.md frontmatter name:; a comment documents why (dedup + disable-set vs. display/activation) and the breakage symptom if they drift.

Verification

• uv run pytest tests/plugins/ tests/test_callbacks_extended.py -q → 2411 passed, 5 skipped
• ruff check + ruff format --check clean on all touched files
• Every load-bearing claim in the SKILL.md (agent discovery precedence, tool registration flow, model selection order, system-prompt layering, skill discovery order) was spot-checked against source.