server: surface unclosed thinking as reasoning_content, not content

[stefanwichmann]

Fixes #509.

Problem

When thinking mode is on and generation stops before </think> (e.g. the response is truncated at max_tokens mid-thought), the non-streaming path in parse_generated_message_ex routed the partial chain-of-thought into content (via split_reasoning_content) with an empty reasoning_content. Clients received raw, unfinished CoT as if it were the assistant's final answer. The live streaming classifier already handles this correctly — only the blocking path was affected.

Fix

In the require_thinking_closed && !think_end branch, treat the accumulated text as unfinished reasoning: surface it as reasoning_content and leave content empty (stripping a leading <think>). Any DSML in that text is unclosed reasoning too, so it's deliberately not executed as a tool call — composing with the existing unclosed-think handling from #318.

Testing

• Adds test_thinking_unclosed_is_reasoning_not_content.
• ./ds4_test --server passes.
• Built with make cuda-spark and validated live on a DGX Spark (GB10): truncated non-streaming responses now return empty content + populated reasoning_content; completed responses unchanged.

Single file, +35/−3.

[Smallfu666]

Independent DGX Spark validation for this PR.

Test environment:

• NVIDIA DGX Spark / GB10 (sm_121)
• Ubuntu 24.04.4 LTS, ARM64
• Kernel 6.17.0-1026-nvidia
• CUDA 13.0
• NVIDIA driver 580.159.03
• DeepSeek V4 Flash Q2 imatrix
• Baseline: main at 80ebbc396aee40eedc1d829222f3362d10fa4c6c
• PR head: 76573b091887d18d2aa3066f7d6da1593191764b

Baseline reproduction:

• A truncated non-streaming thinking response returned finish_reason: "length".
• Partial reasoning was exposed through message.content.
• reasoning_content was empty or absent.

Validation results with this PR:

• make cuda-spark: PASS
• ./ds4_test --server: PASS
• Truncated non-streaming thinking response: PASS — content was empty and partial reasoning was returned through reasoning_content.
• Completed non-streaming thinking response: PASS — final content and reasoning_content remained correctly separated.
• Truncated streaming response: PASS — no reasoning leaked into content.
• No-thinking non-streaming response: PASS — normal assistant output remained in content, without unexpected reasoning_content.
• Unclosed thinking containing DSML-like tool-call text: PASS — no tool call was executed.

The baseline issue was reproducible on this machine, and the tested PR head corrected it without a regression in these five API paths.