[JustVugg]

Fixes #2590.

What was wrong

coli tune rendered every qwen36-family checkpoint's prompts with one template, <|im_start|>user\n{prompt}<|im_end|>\n<|im_start|>assistant\n<think>\n. The family holds checkpoints with different templates (Qwen3.6, Qwen3-Coder, Flint, Muse Spark, LFM3, the dense Qwen3; the gateway tells them apart with detect_chat_flavor). Qwen3-Coder has no thinking mode: after that open <think> it ends its turn on the first token, so the engine did one forward, generated nothing, reported 0 tok/s, and the tune stopped at its baseline. Serving worked because the gateway renders the Coder's own template.

Measured on Qwen3-Coder-30B-A3B int4 (Justvugg/Qwen3-Coder-30B-A3B-colibri-int4) with CUDA on an RTX PRO 4000, the same message as a raw completion, greedy:

prompt	tokens	finish
tune's form (ends in <think>\n)	0	stop
the gateway's rendering (Qwen3-Coder template)	16	length, "Here's a Python function that checks whether a string is a palindrome: ..."

coli tune on dev fails the same way on Linux as in the report on Windows: [chain] qwen36: 1 forwards ... (0 tokens), then tuning failed: baseline: engine reported no decode speed.

Change

• openai_server.tune_chat_prompt(model, text): one user message rendered as the gateway renders a chat request that sets nothing else, with the checkpoint's template and its default reasoning.
• The request handler's defaults are factored out so both paths share them: default_reasoning_effort() (Qwen3.8 and Muse Spark xhigh, MiMo and LFM3 high, COLI_THINK=1, else none) and template_thinks() (false for Qwen3-Coder and OLMoE). The handler behaves as before.
• FamilyDescriptor.tune_renders_chat, true for qwen36 only: coli tune renders that family's prompts with tune_chat_prompt. The other families keep their tune_prompt_template.
• tests/test_tune_chat_prompt.py: the Coder gets its template and no <think> (equal to render_chat_qwen3_coder); Qwen3.6 gets the pre-closed block a chat request gets; COLI_THINK=1 opens it for Qwen3.6 and still not for the Coder; only qwen36 renders through the gateway.

Measured

coli tune on the RTX PRO 4000, CUDA build of dev's engine for both sides (the change touches no engine code):

model	dev	this branch
Qwen3-Coder-30B-A3B int4	fails at the baseline (0 tokens)	runs to the end: baseline, omp-14, omp-7, baseline kept
Qwen3.6-35B-A3B Q4_K_M	runs to the end, baseline 114.2 tok/s	runs to the end, baseline 121.8 tok/s

The two Qwen3.6 numbers are not the same request: dev timed a request that thinks, this branch the one coli serve is sent by default, without thinking (the template's pre-closed block). Gateway, chat template and autotune tests: 543 pass.