test(mcp): assert runner emits tool_call_output_item on MCP failures

[OiPunk]

Summary

This is a follow-up to #2554, adjusted based on maintainer feedback.

Instead of only asserting that the run continues, this test now explicitly verifies that runner-level MCP tool failures are surfaced as a generated tool_call_output_item.

What changed

• Added test_runner_emits_mcp_error_tool_call_output_item in tests/mcp/test_runner_calls_mcp.py.
• Introduced a crashing fake MCP server for deterministic failure behavior.
• Asserted both sync and streaming paths:
  • a tool_call_output_item is produced
  • the item output matches the expected error text from default_tool_error_function
  • the run still continues to the next turn (final_output == "done")

Validation

Locally run:

• env -u all_proxy -u ALL_PROXY -u http_proxy -u HTTP_PROXY -u https_proxy -u HTTPS_PROXY -u no_proxy -u NO_PROXY uv run --with pytest pytest -q tests/mcp/test_runner_calls_mcp.py
• uv run --with ruff ruff check tests/mcp/test_runner_calls_mcp.py
• uv run --with mypy mypy tests/mcp/test_runner_calls_mcp.py

All passed locally.

[seratch]

@OiPunk Thank you for working on this. I think this is worth adding, so I'll merge it.

That said, we're trying to keep review effort focused, so for future test-only PRs, please prioritize cases that cover critical or user-visible paths.