fix(streaming): reconcile pending Responses tool calls on abort

[fscfede-beep]

Summary

Runner-level implementation of streamed Responses abort reconciliation, motivated by openai/openai-python#3561.

When a server-managed Responses stream is cancelled after a function_call has been emitted, the runner records the pending call and reconciles it with a fresh non-streaming model request containing function_call_output(status='incomplete', output='aborted'), then preserves the original cancellation signal.

The generic transport streaming close path is untouched.

Design

• pending calls are tracked in run_internal/run_loop.py, where the runner already owns conversation_id / previous_response_id and turn lifecycle;
• reconciliation uses the existing get_response_with_retry path and server conversation tracker;
• reconciliation is isolated from generic model transport cleanup;
• cancellation is not allowed to cancel the reconciliation task once started;
• terminal response events clear the pending set to avoid duplicate reconciliation.

Tests

Added deterministic tests for early function_call tracking, terminal clearing, and abort payload construction. The repository CI should provide the final verification for the full runner path.

[chatgpt-codex-connector]

You have reached your Codex usage limits for code reviews. You can see your limits in the Codex usage dashboard.