test: add LangGraph node_metrics isolation regression test

[jsonbailey]

Summary

Adds the missing regression test that should have shipped alongside the LangGraph runner migration in #156. Mirrors the OpenAI version added in #155.

Bug class being guarded

The graph-runner pattern uses a callback-style accumulator (LDMetricsCallbackHandler for LangGraph, _node_* state dicts for OpenAI) to collect per-node metrics during a run(). If that accumulator is held on the runner instance across run() calls, the second invocation's node_metrics ends up containing stale entries from the first run.

#155 fixed this for the OpenAI runner and shipped a regression test:

• packages/ai-providers/server-ai-openai/tests/test_openai_agent_graph_runner.py::test_openai_agent_graph_runner_run_resets_node_metrics_between_runs

#156 migrated the LangGraph runner to the same AgentGraphRunnerResult shape but did not include the equivalent regression test. This PR adds it.

What this PR does

Adds test_langgraph_runner_run_resets_node_metrics_between_runs in packages/ai-providers/server-ai-langchain/tests/test_langgraph_agent_graph_runner.py. It calls runner.run() twice on the same LangGraphAgentGraphRunner instance, fires callback events on the handler each time, and asserts that the second result's node_metrics:

• contains only the second invocation's nodes (not a union with the first)
• holds a fresh LDAIMetrics instance, not the same object returned on the first run

No fix to the runner itself was needed: LangGraphAgentGraphRunner.run() already constructs a fresh LDMetricsCallbackHandler per call, so state is isolated by construction. The test locks that behaviour in.

Test plan

• make test-langchain passes (86 tests)
• make test passes across all packages (server-ai 200, langchain 86, openai 44)
• make lint passes (mypy, isort, pycodestyle)
• New test passes both in isolation and as part of the full suite

---

Note

Low Risk
Low risk: test-only change that tightens coverage around metrics collection across repeated run() calls, with no production logic modified.

Overview
Adds a new async regression test in test_langgraph_agent_graph_runner.py that calls LangGraphAgentGraphRunner.run() twice on the same instance and asserts the second result’s metrics.node_metrics and metrics.path are fresh per invocation (no stale/cumulative state from the first run).

The test bypasses graph compilation by injecting a mocked compiled graph whose ainvoke triggers LangChain callback events, specifically guarding against accidental reuse of the metrics callback handler/state across runs.

^{Reviewed by Cursor Bugbot for commit 637ee98. Bugbot is set up for automated code reviews on this repo. Configure here.}