This pull request fixes a streamed output-guardrail regression where Runner.run_streamed() could swallow OutputGuardrailTripwireTriggered until stream_events() was consumed, allowing callers that await run_loop_task directly to observe a completed run with final_output set. It updates the streaming run loop to re-raise output-guardrail tripwires and cancellation, while avoiding the generic agent-span error path so the existing guardrail-specific tracing remains intact.

It also adds a regression test that exercises the non-stream_events() path by awaiting run_loop_task directly and asserting that the tripwire is surfaced immediately and the final output is not finalized. This keeps the fix aligned with the real repro instead of only covering the already-tested event-consumption path.