fix(llm): keep finish_reason and empty content for tool responses without calls

[MohammadHijjawi97]

What

ChatModelMixin._post_process_completion (non-streaming chat with tools, used by the vLLM, SGLang, MLX, llama.cpp and transformers engines) has two problems when the model does not call a tool:

• finish_reason is always "stop", so a response cut off by max_tokens is reported as "stop" instead of "length". The streaming tool path and the request without tools keep "length".
• When the output only contains reasoning (e.g. max_tokens is reached inside the thinking block), the reasoning parser returns empty content, but the if processed_content: check then keeps the raw output, so the whole reasoning is returned again as content.

Example with Qwen3 (reasoning_content=True), output cut off while thinking:

Before:

with tools   : {"content": "<think>\nThe user wants the weather, I should call", "tool_calls": [], "reasoning_content": "<think>\nThe user wants the weather, I should call"}, "finish_reason": "stop"
without tools: {"content": "", "reasoning_content": "<think>\nThe user wants the weather, I should call"}, "finish_reason": "length"

After:

with tools   : {"content": "", "tool_calls": [], "reasoning_content": "<think>\nThe user wants the weather, I should call"}, "finish_reason": "length"

Fix

Always use the reasoning parser's content for tool parsing, and use the engine's finish_reason when there is no tool call ("stop" if it is missing). Responses with tool calls are unchanged.

Tests

• Added test_post_process_completion_stopped_while_thinking (fails before the fix, passes after).
• pytest xinference/model/llm/tests/test_utils.py xinference/model/llm/tool_parsers/tests passes.
• black / ruff / isort / codespell / mypy (pre-commit) clean on changed files.

[qinxuye]

LGTM. Verified the reasoning-only empty-content handling and preservation of the engine finish_reason when no tool is called. All 176 focused utility and tool-parser tests pass locally. The sole CI failure is test_fatal_engine_error_closes_model in the image engine, outside the changed LLM post-processing path.