Non-streaming responses leak <think> tags into content for qwen3.5, qwen3.6-moe, qwen3vl and qwen3.5-omni

Summary

On the non-streaming OpenAI path, models whose parse_nstream_content does not
split the think block deliver the reasoning inside content instead of
reasoning_content, so the client shows raw <think> tags as the answer. When
generation is truncated mid-thought, the partial thought becomes the entire answer.

Affects qwen3.5, qwen3.6-moe, qwen3vl and qwen3.5-omni — all
reasoning-capable. Reproduced on main at 92f3f13.

Environment

• OS: Ubuntu 26.04.1 LTS (kernel 7.0.0-31-generic)
• NPU driver: NPU FW 1.1.2.64, amdxdna 0.7
• flm version: FLM v1.0.6
• flm validate:

[Linux]  Kernel: 7.0.0-31-generic
[Linux]  NPU: /dev/accel/accel0 with 8 columns
[Linux]  NPU FW Version: 1.1.2.64
[Linux]  amdxdna version: 0.7
[Linux]  Memlock Limit: infinity

Reproduce

flm serve qwen3.5:9b
curl -s localhost:52625/v1/chat/completions -H 'Content-Type: application/json' -d '{
  "model": "qwen3.5:9b",
  "messages": [{"role":"user","content":"What is 5 plus 6? Think briefly."}],
  "max_tokens": 400, "think": true
}' | jq '.choices[0].message'

Observed on 92f3f13 — no reasoning_content key, thought in content:

finish=stop    reasoning_content=False  <think>_in_content=True
content: '<think>\nThinking Process:\n\n1.  **Analyze the Request:**\n    *   '

Same with "max_tokens": 60 (finish=length), where the truncated thought is
the whole answer.

Cause

build_nstream_response takes reasoning_content from parse_nstream_content,
and only some models populate it:

The streaming path is unaffected: its parser switches to REASONING on
<think> and stays there, so streamed responses split correctly. Only the
non-streaming path is wrong.

Impact

Agent clients render the thought as the assistant's reply. A truncated thought
reads as the prompt being echoed back ("Thinking Process: 1. Analyze the
Request: The user wants me to ..."), which looks like the model looping.

Fix

PR follows: split the block once in build_nstream_response so every model
reports reasoning the same way, and route an unclosed (truncated) block to
reasoning_content with empty content. No-op for models that already
populate it.

Start by locating build_nstream_response and parse_nstream_content, then run the provided flm serve and curl reproduction for qwen3.5 with both max_tokens values. Compare non-streaming output with the streaming behavior described in the issue. Done means reasoning is reported in reasoning_content, content is empty for a truncated thought, and models that already split reasoning remain unchanged.