feat(ai): LangChain agent test harness — fake() AIMessage support, fake_chat_model, Agent.record()

[bedus-creation]

What

Adds a lightweight test harness on top of LangChain so tests can run LLM agents without hitting a real model API. Builds on the existing `ai/` module (whose `Agent.fake()` glob-stubbing + `AgentSnapshot` replay already existed) and fills in the LangChain pieces sketched in the reference prototype (`example/agents/tinker.py`).

This is the lower-risk, additive approach: the existing raw-SDK `_run` path is untouched, and all prior tests stay green.

Changes

• `fake()` accepts LangChain `AIMessage` values (alongside `AgentResponse` / `AgentSnapshot`). On a pattern match the `AIMessage` is converted to an `AgentResponse`, preserving `content` and `tool_calls` — the provider is never called.
• `fake_chat_model(turns)` / `Agent.fake_model(turns)` — build a fake chat model that replays a scripted sequence of assistant turns (`AIMessage` or `str`). Pass it straight to `langchain.agents.create_agent(model=..., tools=[...])` to drive a full tool-calling loop offline (`human → ai(tool call) → tool → ai(final)`). Includes the `bind_tools()` shim that `GenericFakeChatModel` lacks (otherwise `create_agent` raises `NotImplementedError` when tools are present).
• `Agent.record(path, , pattern="")` — explicit VCR-style record-then-replay entry point. First run calls the real provider and saves the response; later runs replay from disk.
• New optional `[langchain]` extra (`langchain>=1.0`, `langchain-core>=1.0`). All langchain imports are lazy, so the base package imports without the extra installed.

Design notes

• LangChain's `ToolCall` is a `TypedDict` (unhashable), so it cannot be a literal dict key — the prototype's `{ToolCall(...): AIMessage(...)}` would raise `TypeError` at construction. Tool-call sequencing is therefore expressed the idiomatic LangChain way: an ordered list of turns fed to `fake_chat_model`, where tool-calling turns are `AIMessage(content="", tool_calls=[...])`.

Tests

• `tests/ai/test_agent_record.py` — `record()` VCR semantics (replay when cassette exists, record+save when missing, custom pattern, chaining, roundtrip).
• `tests/ai/test_agent_langchain.py` — `AIMessage` fakes (content + tool_calls, no `_run`), and a real `create_agent` tool loop running entirely offline. Guarded with `importorskip` so envs without the extra skip cleanly.
• Full suite: 1559 passed, 7 skipped; ruff clean.