fix: apply Claude thinking parameters from the evaluated LangChain config

[andrewklatzke]

Summary

• Stop hardcoding Anthropic thinking budget and max tokens in the LangChain thinking example so flag model.parameters are applied after evaluation.
• Align LangChain README factory examples so the model name wins over a colliding model key in the parameter bag.
• Refresh uv.lock workspace versions to match the published package metadata.

Fixes AIC-3400.

Test plan

• uv run python main.py langchain-thinking launch-darkly-documentation-summarizer-messages-claude "Reason it out yourself without any tools: what is 17 times 23?" exits 0 with a non-empty response and usage totals > 0
• Same example with doesnt-exist exits 1 with a clean error and no new output/ JSON
• uv run python main.py native-graph-langchain travel-agent-flow "Book me a flight to Paris" exits 0 with a non-empty response and usage totals > 0
• Same example with travel-agent-flow-wrong-key exits 1 with a clean disabled-graph error and no new JSON

---

Note

Overview
Stops overriding Anthropic extended-thinking settings in the LangChain thinking example so ChatAnthropic is built from the evaluated flag’s model.parameters (plus the configured model name), instead of hardcoded thinking / max_tokens constants.

Updates LangChain agents and messages README factory snippets so the flag’s model name is applied after spreading model.parameters, ensuring it wins if parameters include a colliding model key.

Bumps workspace package versions in uv.lock (e.g. 0.2.1 → 0.2.2) to align with published metadata.

^{Reviewed by Cursor Bugbot for commit 408ad5b. Bugbot is set up for automated code reviews on this repo. Configure here.}