fix: cancel RunResultStreaming on retry to prevent leaked asyncio tasks

When stream_events() is interrupted by a timeout and we retry, the old
RunResultStreaming object's _run_impl_task continues running as a dangling
asyncio.Task. Calling aclose() on the stream_events() async generator
throws GeneratorExit which skips the SDK's _cleanup_tasks() call, so
the background task is never cancelled.

Each timeout retry leaked one task still streaming from the API and
holding MCP server references. On process exit, asyncio.run() blocks
trying to clean up these orphaned tasks — a likely contributor to the
post-completion hangs requiring watchdog intervention.

Fix: call result.cancel() in the finally block after aclose(). This
cancels _run_impl_task, input/output guardrail tasks, and clears the
event queue before retrying with a fresh RunResultStreaming.

Author: anticomputer

src/seclab_taskflow_agent/runner.py

@@ -436,6 +436,7 @@ async def _run_streamed() -> None:
                 max_retry = MAX_API_RETRY
                 rate_limit_backoff = RATE_LIMIT_BACKOFF
                 while rate_limit_backoff:
+                    result = None
                     try:
                         result = agent0.run_streamed(prompt, max_turns=max_turns)
                         stream = None
@@ -464,6 +465,12 @@ async def _run_streamed() -> None:
                                 aclose = getattr(stream, "aclose", None)
                                 if aclose is not None:
                                     await aclose()
+                            # Cancel the RunResultStreaming background tasks.
+                            # aclose() on the stream_events() async generator throws
+                            # GeneratorExit which skips _cleanup_tasks(), so we must
+                            # cancel explicitly to avoid leaking _run_impl_task.
+                            if result is not None:
+                                result.cancel()
                         await render_model_output("\n\n", async_task=async_task, task_id=task_id)
                         return
                     except APITimeoutError: