test(app): route prompt e2e through mock llm

[kitlangton]

Summary

• extend the shared TestLLMServer to serve OpenAI Responses API streams so app e2e can reuse the same mock LLM abstraction
• switch the prompt reply e2e to update backend config through the supported global config API and assert the mock server is actually called
• keep the active model on a built-in OpenAI model while routing its transport to the mock server, including the extra queued response used by the title request