feat: Migrate LangGraph runner to AgentGraphRunnerResult; clean up legacy shape detection

[jsonbailey]

Summary

• Updates LangGraphAgentGraphRunner.run() to return AgentGraphRunnerResult (with GraphMetrics) instead of the legacy AgentGraphResult shape
• Adds collect_node_metrics() to LDMetricsCallbackHandler for pure data extraction (no LD tracker calls from the runner)
• Removes the transitional isinstance(raw_result, AgentGraphRunnerResult) detection branch from ManagedAgentGraph — now that both OpenAI and LangGraph runners return the new shape, the bridge is dead code
• All graph-level and per-node tracking events are driven exclusively by ManagedAgentGraph._flush_graph_tracking()
• Updates test_langgraph_agent_graph_runner.py to verify runner returns AgentGraphRunnerResult and does NOT call graph tracker directly
• Updates test_tracking_langgraph.py to verify events via handler flush() or result shape; removes direct runner → graph tracker assertions

Stack position

This is PR 11-langchain, stacked on:

• PR feat: Add ManagedGraphResult, GraphMetricSummary, and AgentGraphRunnerResult types #151 (ManagedGraphResult types)
• PR feat: Graph tracking refactor — ManagedAgentGraph drives tracking for new runner shape #154 (graph tracking refactor — transitional bridge)
• PR feat: Update OpenAI graph runner to return AgentGraphRunnerResult with GraphMetrics #155 (OpenAI runner returns AgentGraphRunnerResult)

After this PR, the legacy AgentGraphResult shape is fully retired from production code.

Test plan

• make test passes in packages/ai-providers/server-ai-langchain (81 tests)
• uv run pytest packages/sdk/server-ai/tests/ passes (149 tests)
• make test passes in packages/ai-providers/server-ai-openai (40 tests)

🤖 Generated with Claude Code

---

Note

Medium Risk
Changes the agent-graph runner return type and shifts graph/per-node tracking responsibility to ManagedAgentGraph, which can affect downstream integrations expecting the legacy AgentGraphResult shape or runner-driven tracking behavior.

Overview
Migrates LangGraphAgentGraphRunner.run() to return AgentGraphRunnerResult with GraphMetrics (including path, duration, total token usage, and per-node LDAIMetrics) instead of the legacy AgentGraphResult, and removes runner-side graph tracker calls and per-run eval-task plumbing.

Simplifies ManagedAgentGraph to require the new runner result shape and to emit all graph-level and per-node LaunchDarkly tracking events from result.metrics.node_metrics.

Updates the LangChain callback handler to deprecate flush() for production tracking and adds a pure collect_node_metrics() extractor, plus adjusts tests to assert the new result fields (content, duration_ms, path) and to verify tracking now occurs via the managed layer (including running some integration tests through ManagedAgentGraph).

^{Reviewed by Cursor Bugbot for commit 0daafb6. Bugbot is set up for automated code reviews on this repo. Configure here.}