fix(langchain): add provider metadata, route tools to metadata, allowlist span fields

[starfolkai]

Summary

Align the LangChain callback handler with .agents/skills/sdk-integrations/SKILL.md. Audit turned up five issues:

• Missing metadata.provider on every span — spec requires every llm span to carry both metadata.model and metadata.provider. Derive provider at start from serialized["id"] (strip langchain_ prefix + small alias table for google/aws/azure/mistral) and at end from response_metadata.model_provider as a fallback.
• Tools in invocation_params instead of metadata.tools — spec says tool definitions live in metadata.tools. Lift tools / functions to metadata.tools and tool_choice / function_call to metadata.tool_choice via _split_tools, stripping them from the invocation params that flow into metadata.
• Denylist → allowlist — dropped the exclude_metadata_props regex (dead code — was set but never applied) and stopped spreading **kwargs and the full serialized runnable dump into span metadata across every callback. Each callback now captures only LangChain user metadata (preserved as metadata.metadata so deepagents tests still resolve ls_integration / lc_agent_name / lc_versions), tags, and known-safe fields.
• Excess serialization — on_tool_start no longer duplicates the tool payload as inputs + input_str + input (parsed) + serialized; keeps only the parsed input. Removed the raw kwargs dump in every error callback.
• on_agent_action span type — LLM → TOOL (the callback fires when the agent picks a tool, not when it calls an LLM).

Also trimmed narrator comments per the skill's "default to no comments" guidance, and merged the two independent _walk_generations passes in on_llm_end into a single _model_and_provider_from_response walk.

The LangChain-native message / LLMResult shape in input / output is left as-is — the Braintrust backend already normalizes it.

Test plan

• Extended existing VCR-backed tests to assert metadata.provider ("openai" for ChatOpenAI spans, "anthropic" for ChatAnthropic), and metadata.tools presence for the test_tool_usage case.
• cd py && CI=1 BRAINTRUST_TEST_PACKAGE_VERSION=latest pytest src/braintrust/integrations/langchain/ — 18 passed, 1 skipped.
• cd py && CI=1 BRAINTRUST_TEST_PACKAGE_VERSION=0.3.28 pytest src/braintrust/integrations/langchain/test_callbacks.py src/braintrust/integrations/langchain/test_anthropic.py — 13 passed, 1 skipped.
• pylint --errors-only clean.
• No cassettes re-recorded, no new mocks/fakes — used the existing VCR-backed suite.

🤖 Generated with Claude Code

Created by abhijeet

Slack thread