Bug: openai_codex uses a per-turn prompt hash as prompt_cache_key instead of a stable conversation key