feat(server): add --default-thinking-mode flag for server-wide thinking default

[rafaeldrincon]

Problem

OpenAI-compatible clients that never send chat_template_kwargs (Vercel AI SDK openai-compatible provider, llama-swap, LiteLLM proxies) get the model's template default, which for reasoning models like Qwen3.6 means thinking is always on. The token budget goes to reasoning_content and content comes back empty.

Measured on real hardware: a 200-token max_tokens budget produced 199 reasoning tokens and 0 content tokens; the client saw an empty answer.

Current workaround (--reasoning-parser off) just dumps thinking into content — still wasting the budget.

Solution

--default-thinking-mode {auto,chat,thinking} server flag:

• auto (default): current behavior, the template decides
• chat: fills enable_thinking=False into every request that does not set any explicit thinking control
• thinking: fills enable_thinking=True the same way

An explicit per-request value (enable_thinking / thinking / thinking_mode in chat_template_kwargs, or the DeepSeek thinking.type wire toggle) always wins over the server default — the flag only fills what the request left unset.

This mirrors llama.cpp's --jinja template behavior where the server owns the default and clients can override per-request.

Implementation

• ServerArgs.default_thinking_mode + argparse entry (server/args.py)
• apply_default_thinking_mode() merge helper in server/openai_api.py
• Applied at all three frontends: OpenAI chat completions, Anthropic messages, Responses API
• 6 unit tests in tests/server/test_default_thinking_mode.py (verified all pass)

Closes #472

Test environment

ft version: 0.1.2
model: nvidia/Qwen3.6-35B-A3B-NVFP4
gpu: RTX 3080 Ti Laptop (16GB, sm_86)
cuda: 13.1, driver: 580

With --default-thinking-mode chat the same Understory agent workload that previously returned empty content produces correct answers with no client changes.