Python: AgentLoopMiddleware does not close the active inner stream when streaming is stopped early #9328

Describe the bug

When a streaming AgentLoopMiddleware run is closed before the active model stream is exhausted, the middleware closes its outer async generator but does not close the inner ResponseStream currently being consumed. As a result, the inner stream's cleanup hooks do not run deterministically. Resources owned by the provider stream, such as HTTP response bodies, telemetry scopes, or other transport cleanup, can remain live until garbage collection.

The non-loop streaming path and ResponseStream wrappers propagate close() to their active inner stream. The loop middleware inserts an async-generator layer (async for update in inner: yield update) without retaining and closing inner in its finally block, so that propagation stops at the loop boundary.

To reproduce

1. Use a chat client whose streaming ResponseStream has a cleanup hook.
2. Run an Agent with AgentLoopMiddleware and stream=True.
3. Consume one update and call await stream.close() before the provider stream is exhausted.
4. Observe that the provider stream cleanup hook has not run.

A minimal regression assertion is:

stream = agent.run("start", stream=True)
await anext(stream)
await stream.close()
assert cleanup_calls == 1  # currently 0

Expected behavior

Closing or cancelling the outer loop stream should close the active inner ResponseStream exactly once and run its cleanup immediately. Normal full consumption should remain unchanged.

Environment

• Current main (b53d25ed1cd2f54ee5e5e5ed548bb2c5515022f3)
• Python 3.13.15
• macOS