[mason] Fix agent-langgraph template dropping LangGraph/tool spans from traces

[jamesbxwu]

Problem

The agent-langgraph template produces malformed MLflow traces: the trace shows only the runtime's per-request root span, with the entire LangGraph subtree - model calls, the tools node, and the individual tool spans - missing. Tool calls never appear grouped under the root span.

Root cause

agent/agent.py imported configure_tracing from the top-level databricks_mason:

from databricks_mason import configure_tracing, ...

That resolves to the framework-neutral runtime.tracing.configure_tracing(autolog=None). With no autolog bound, it enables tracing (a destination + experiment are configured) but never calls mlflow.langchain.autolog(). runtime.py still opens its manual per-request span, so a trace is produced - but nothing emits the LangGraph/LangChain spans, leaving a single-span trace.

The sibling agent-openai template already imports configure_tracing from databricks_mason.openai (which binds mlflow.openai.autolog). This brings the langgraph template in line by importing from databricks_mason.langgraph (which binds mlflow.langchain.autolog).

Fix

• Import configure_tracing from databricks_mason.langgraph instead of the neutral top-level one.
• Add a hermetic regression test asserting configure_tracing() enables LangChain autologging when tracing is configured. It fails on the old import and passes on the fix.

Verification (end-to-end, live workspace)

Ran the actual template server against a live workspace with MLflow tracing to a real experiment and a real tool-calling model.

Before: trace had a single span:

stream_handler

After: full nested tree:

stream_handler
  └ LangGraph
      ├ model → _RoutedChatDatabricks
      ├ tools → get_current_time      ← tool call now under the root
      └ model → _RoutedChatDatabricks

Template test suite: 16 passed, 1 skipped (the live model test).

This pull request and its description were written by Isaac.