[NVolcz]

The server applies a request's settings (think, reasoning_effort, temperature, top_p, top_k, min_p, the penalties) to the shared engine and never resets them. A setting one request gives stays in place for every later request that leaves it out, from any client. For example, after one request with "think": true, requests without think still render <|think|> and the model keeps reasoning. The docs warned about this ("Settings stick").

The engine now remembers its sampler config and whether thinking is on once the model is loaded, and the server puts them back before applying each request's settings, on the chat endpoints and on /api/generate. A setting a request leaves out falls back to the model's default. Because the sampler is recreated, every request also starts with an empty penalty window.

The docs, a code comment and the API tests' notes now describe the new behaviour, and test_reasoning.think_does_not_carry_over checks it.

Qwen3_5_Omni::set_sampler did the same as AutoModel::set_sampler, so it is removed; otherwise Qwen3.5-Omni would skip remembering its sampler config.

The reasoning_effort levels of GPT-OSS and Hy-Memory, conversational_mode and image-max-tokens are stored in model-specific fields and still stay in effect until the model is reloaded.

I noticed this while reading the prompts flm renders, with a small debug branch that logs every prompt and raw output to a file (log-prompt, FLM_LOG_PROMPT=<file>): a request without think still had <|think|> in its prompt.