The chat_template shipped here was byte-identical to Qwen2.5's, so <think>
blocks were never stripped from earlier assistant turns, enable_thinking was
unavailable and a Qwen2.5 default system prompt was injected.
This is Qwen3's upstream template with the last_query_index prescan removed, so
reasoning is stripped from every prior assistant turn unconditionally. That
keeps rendering monotonic, which append-only KV cache prefilling requires.
Update chat template for strict monotonicity and tool support
