qwen3_coder non-streaming drops the text between tool calls #89

FunctionCallParser.parse_non_stream keeps only the text before the first <tool_call> (stripped) plus the text after the last </tool_call>. Prose between two calls disappears from non-streaming content, while streaming emits it. Streaming's own whitespace rule (suppress_ws) also depends on chunking: a whitespace-only gap after a call is dropped only when it arrives as a separate chunk. An empty <tool_call></tool_call> block is hidden by streaming but left in non-streaming content.

from freetoken.server.function_call_parser import FunctionCallParser
tools = [{"type": "function", "function": {"name": "read", "parameters": {"type": "object", "properties": {"p": {"type": "string"}}}}}]
blk = lambda v: f"<tool_call>\n<function=read>\n<parameter=p>\n{v}\n</parameter>\n</function>\n</tool_call>"
r = FunctionCallParser(tools, "qwen3_coder").parse_non_stream("Hi" + blk("a") + "\nThen more.\n" + blk("b"))
print(repr(r.normal_text), len(r.calls))   # 'Hi' 2  (streaming shows 'Hi\nThen more.\n')

Fixed on 5080deploy by 7339156 (merged in 1542d52); stream/non-stream parity tests are in tests/server/test_streaming_model_matrix.py.