self-assess whether the change is an instance of a broader pattern that warrants its own generalization card. Design A+B+A: lightweight prompt-only / code-mutating-turns only / reminder-only.
(max_query_len=1), following the same pattern used by vLLM's EAGLE speculative decoding proposer. Registered via the ``vllm.general_plugins`` entry point so it is auto-discovered by all vLLM processes ...