llama-server chokes on spaced kwargs

This looks like a brittle config-parsing footgun in `llama-server`, not a Qwen3.6 model behavior issue. If the parser treats whitespace as significant, it can make identical configs behave differently and waste a lot of debugging time.

• –The failure mode is tiny but nasty: `{ "preserve_thinking": true }` breaks, while `{"preserve_thinking": true}` works.
• –That points to server-side config handling, so users may incorrectly blame the model, template, or reasoning settings.
• –For self-hosted AI stacks, whitespace-sensitive JSON in INI files is a reproducibility hazard because copied configs stop being portable.
• –Qwen3.6 reasoning controls are already easy to mix up, so exact documented examples and parser tests matter here.
• –This is most relevant to operators running local inference servers, not end users of hosted chat apps.