Add Guardrail Sandwich LangGraph + LangChain reference implementations

[webup]

Summary

• Add LangGraph StateGraph and LangChain middleware tutorial notebooks for the Guardrail Sandwich pattern — the first of 28 planned reference implementations
• Add shared infrastructure: model_config.py (ernie provider via langchain-dev-utils), hooks.py (shared hook factories), .env.example, REFERENCE_IMPL.md
• Both notebooks are committed with executed outputs, HTML exports, and Markdown exports for GitHub viewing

What's in each commit

1. Project setup — pyproject.toml deps (langgraph optional group + dev), .env.example with AI Studio default, model_config.py with register_model_provider + load_chat_model
2. Shared hooks — hooks.py with HookResult/HookPhase enums + 3 hook factories (amount_threshold, blocklist, output_schema), imported by both notebooks
3. LangGraph notebook — StateGraph with pre_hooks/execute_tool/post_hooks nodes, conditional routing, ToolNode, 4 mock scenarios + real LLM run, graph visualization
4. LangChain notebook — @wrap_tool_call decorator + GuardrailSandwichMiddleware class, both tested with create_agent + live ernie-5.1, agent graph visualization
5. REFERENCE_IMPL.md — overview doc with implementation status, 8-wave roadmap, coming-soon items (Agent Chat UI, Agent Service Runtime)
6. CI fix — drop Python 3.10 from matrix to match requires-python >= 3.11

Design decisions

• Notebooks as the primary artifact — no separate pattern.py for the framework implementations; everything inline for teaching
• Shared hook factories — hooks.py ensures both notebooks use identical hook definitions
• Official APIs only — LangChain notebook uses wrap_tool_call, AgentMiddleware, create_agent from langchain.agents.middleware; no homebrew types
• langchain-dev-utils for model registration — register_model_provider("ernie", "openai-compatible", base_url=...) + load_chat_model("ernie:ernie-5.1")
• Hook design aligned — same factory signatures, same config dict shape, same scenarios across both notebooks

Test plan

• CI passes — Python 3.11 + 3.12 (pure-Python tests + examples)
• uv sync --extra langgraph --extra dev succeeds
• uv run pytest --nbmake --nbmake-timeout=120 action/d-guardrail-sandwich/langgraph/tutorial.ipynb action/d-guardrail-sandwich/langchain/tutorial.ipynb — both pass
• uv run pytest --import-mode=importlib -v — existing pure-Python tests still pass
• HTML exports render correctly in browser
• Markdown exports render on GitHub

[huangjia2019]

Excellent work, Haili — and an unusually strong opener for a 28-pattern series.

A few things I want to put on record about this PR, since it sets the bar for the remaining 27:

• Shared hooks.py + model_config.py is the right architectural call. Keeping the hook factories identical across the LangGraph and LangChain notebooks means readers can compare frameworks without wondering whether differences come from the pattern or from implementation noise.
• Dual-framework reference (StateGraph + middleware) in one folder is the teaching format I was hoping for. The visible graph vs. middleware-plugged-into-create_agent contrast does a lot of work for the reader.
• Official APIs only — no homebrew types is exactly the standard ADPS needs to anchor against. Anyone reading this should be able to map directly to upstream LangChain / LangGraph docs.
• The 8-wave roadmap + the Agent Chat UI / Agent Service Runtime "coming soon" note give the repo a future trajectory beyond static reference code. Both very welcome.

CI green on 3.11 / 3.12, mergeable clean. Squashing and merging now. Looking forward to Wave 1.

—— 顺便：感谢海立老师亲自给 ADPS 第一份代码 reference 立下这个标杆。这一 PR 会作为 ADPS-CORE 资产收录进 Master 控制板。