Add LangChain callback handler integration

[crtahlin]

Motivation

LangChain is the most popular framework for building LLM-powered agents. Adding a provenance callback handler would let LangChain users add audit trails to their agents with a single line of code:

from datafund_provenance.integrations.langchain import ProvenanceCallbackHandler

handler = ProvenanceCallbackHandler(gateway_url="https://provenance-gateway.datafund.io")
llm = ChatOpenAI(callbacks=[handler])

This unlocks:

• "One-line audit trails for LangChain agents" narrative (blog post + co-marketing with LangChain team)
• Discovery via LangChain's integrations docs and community
• Natural entry point for developers already building agents

Part of the fellowship marketing plan (Month 2, Week 5-6).

Requirements

Callback Handler

• Implement ProvenanceCallbackHandler extending LangChain's BaseCallbackHandler
• Log key events to provenance gateway:
  • on_llm_start / on_llm_end — capture prompts, completions, model info, token usage
  • on_chain_start / on_chain_end — capture chain execution with inputs/outputs
  • on_tool_start / on_tool_end — capture tool calls and results
  • on_agent_action / on_agent_finish — capture agent reasoning steps
• Each event uploads a provenance record (JSON metadata → gateway → Swarm)
• Support session/run grouping (link related events via run_id)
• Optional notary signing for each record
• Configurable: which events to log, verbosity level, async vs sync

Packaging

• Add as optional dependency group: pip install swarm-provenance-cli[langchain]
• LangChain should be a lazy import (don't break CLI for users without langchain installed)
• Place under swarm_provenance_uploader/integrations/langchain.py

Testing

• Unit tests with mocked gateway (no real uploads)
• Integration test with real gateway (marked @pytest.mark.integration)
• Example script demonstrating usage

Documentation

• README section on LangChain integration
• Example in examples/ directory

Reference

• LangChain callback docs: https://python.langchain.com/docs/how_to/custom_callbacks/
• Existing gateway client: swarm_provenance_uploader/core/gateway_client.py
• The handler should use the existing GatewayClient for uploads