Investigate frozen LLM fixture replay bypass in hello-world cross test

Summary

tests/cross/test_hello_world.py appears to load real frozen LLM completion fixtures via create_real_llm_responses_from_fixtures(fncall_raw_logs), but then immediately replaces them with synthetic mock responses before running the main hello-world flow.

real_responses = self.create_real_llm_responses_from_fixtures(fncall_raw_logs)

# Always use mock responses for consistent behavior
# Real fixture data may have different tool call sequences than current agent
real_responses = self.create_mock_llm_responses()

That makes the test look like a real frozen-trajectory replay test while mostly exercising hand-written mocks. We should revisit whether these fixtures can become a first-class deterministic replay suite for real LLM trajectories, or rename/split the current test so the behavior is clearer.

This issue was created by an AI agent (OpenHands) on behalf of the user.

Start in tests/cross/test_hello_world.py and trace create_real_llm_responses_from_fixtures(fncall_raw_logs) through the hello-world flow, then compare it with create_mock_llm_responses(). Determine whether the frozen fixtures can support deterministic real-trajectory replay; otherwise make the test naming or split clearly distinguish fixture replay from mock behavior.