Point vLLM (or any OpenAI-client-based consumer) at Algenta's own OpenAI-compatible /v1/chat/completions and /v1/responses surface -- a thin base-url/api-key helper plus a real, honest capability-fidelity conformance suite