pi-neuralwatt v0.16.2

4bcbe4c: Fix reasoning replay on the openai-completions surface: prior-turn chain-of-thought is replayed in the reasoning field (the field K3 streams), but the served chat templates only render reasoning_content, so the model never saw its earlier thinking on any turn after the first. The surface now renames reasoning to reasoning_content on replayed assistant messages. This is a move, not a copy — an empty reasoning_content next to a populated reasoning is preferred by the gateway and would drop the replay again.
b94900e: Update the qwen3.6-35b family (standard, fast, flex) context window and max output tokens in the offline fallback table from 131056 to 262128 to match the live Neuralwatt API.