llama-server marks truncated replies and tool calls as completed

In llama.cpp build b11753, llama-server's /v1/responses endpoint returns status "completed" and incomplete_details: null even when a reply is cut off by max_output_tokens. A truncated tool call arrives with invalid JSON arguments but is also marked completed, while /v1/chat/completions correctly reports finish_reason: "length". No upstream fix exists yet.

• A request with max_output_tokens: 8 returned "1 2 3 4" with status completed
• A write_file call cut off at '{"path": "/' was marked completed
• The stream lacks response.incomplete and function_call_arguments.done events
• The only truncation signal is usage.output_tokens equal to the limit