feat(llm): let cloud providers pass vendor-specific body fields

[sosidudku1]

Problem

buildOpenAiChatBody assembles a fixed set of OpenAI schema fields, so a deployment cannot reach vendor extensions that live outside that schema.

Concretely: Alibaba Model Studio gates reasoning through chat_template_kwargs.enable_thinking. There is currently no way to send it short of adding a bespoke field per vendor.

Change

Optional extraBody object on openai-compatible and qwen-openai-compatible provider entries, merged into the chat completion body.

{
  "id": "model-studio",
  "kind": "qwen-openai-compatible",
  "baseUrl": "https://dashscope-intl.aliyuncs.com/compatible-mode",
  "defaultChatModel": "qwen3.8-27b",
  "extraBody": { "chat_template_kwargs": { "enable_thinking": false } }
}

model, messages, stream and tools are re-applied after the merge. A config entry cannot detach the request from the resolved model, rewrite the prompt, flip streaming, or smuggle in a tool contract the caller never set (the last case is covered by a test).

Scope notes

• Local (llama-server) requests are untouched. Tool calling there runs through GBNF, and sampling-filter already strips grammar before any cloud call, so the two paths never meet.
• With extraBody absent the body is byte-identical to before, asserted by a test.
• openrouter, aimlapi and gemini are deliberately left alone. They can opt in later if a vendor knob calls for it.
• Config parsing rejects a non-object extraBody rather than passing it through.

Testing

• 6 new unit tests (4 body builder, 2 config parser).
• Full suite: 6 failing files, identical set to origin/main at the time of writing (TUI / sidecar / fs-glob). No regressions introduced.

Motivation

Measuring Qwen thinking modes (full / capped / off) through the agent rather than a side script, so every arm of the comparison runs under identical conditions.