qwen3 whitespace around </think> differs between streaming and non-streaming #88

Non-streaming strips content after </think> at both ends, and strips reasoning at both ends. Streaming emits the "\n\n" that follows </think> as a content delta, and emits the newline before </think> as reasoning whenever that newline arrives as its own token. So joined streamed content is "\n\nAnswer" where non-streaming returns "Answer".

p = ReasoningParser("qwen3", force_reasoning=True); out = [p.parse_stream_chunk(t) for t in ["r", "\n", "</think>", "\n\n", "Answer"]]
print(out)                                                                            # ('\n', '') ... ('', '\n\n')
print(ReasoningParser("qwen3", force_reasoning=True).parse_non_stream("r\n</think>\n\nAnswer"))  # ('r\n', 'Answer') before generation's strip

Fixed on 5080deploy by 362719c (merged in 1542d52); stream/non-stream parity tests are in tests/server/test_streaming_model_matrix.py.
Seen live on the previous deploy (8eb353b): with thinking on, every streamed answer and tool-call turn starts with a "\n\n" content delta that the non-streaming response does not have (3 of 3 thinking-on turns in a greedy stream-vs-non-stream check).