fix: avoid content leak when generation ends inside thinking block

[mlpy0]

When generation ends while the parser is still inside <think>...</think>, the chunk that carries finish_reason was unconditionally stamped is_thinking=False in parse_thinking_models. Its text then routed to the content channel, leaking 1-8 characters of the last thinking token into content while the rest of the thinking output stayed in reasoning_content.

This change preserves is_thinking on the final token's text via a separate chunk (with finish_reason=None) and emits an empty-text content chunk to carry finish_reason. Consumers that read only content and finish_reason still see the terminating delta; consumers that read reasoning_content get the full thinking output.

Reproduction (Qwen3.6-27B-8bit, max_tokens=10, default thinking):

• before: usage.completion_tokens=10, content=' '
• after: usage.completion_tokens=10, content=None, reasoning_content populated

Tests: strengthens TestThinkingModelsFinishReason.test_finish_reason_during_thinking to pin the leak, and adds test_finish_reason_during_thinking_no_content_leak covering the starts_in_thinking=True path that hit the bug in production.

[Evanev7]

sounds like an issue in a downstream client or in our adapters. can you tell us what tools you were using that caused this output, and which endpoint (ollama, responses, chat-completions or claude)?

[mlpy0]

Endpoint: chat-completions (POST /v1/chat/completions, stream=false).

Client: small Python script using urllib.request from the stdlib. No SDK, no wrapper, no adapter on top. Just reads choices[0].message.content out of the JSON.

Model: mlx-community/Qwen3.6-27B-8bit (also reproduced intermittently on Qwen3.6-35B-A3B-8bit).

Sample request that reproduced it:

{
  "model": "mlx-community/Qwen3.6-27B-8bit",
  "messages": [{"role": "user", "content": "Count from 1 to 100, comma separated."}],
  "max_tokens": 10,
  "temperature": 0.0,
  "stream": false
}

Response had completion_tokens=10, finish_reason="stop", content=" ". Sweeping max_tokens up to 1000 keeps the same shape: completion_tokens tracks max_tokens, content stays 1 to 8 characters of whatever the last mid-thinking token decoded to.

Since the script does nothing beyond reading the content field out of the JSON, no downstream parser is in play on this side.