/v1/responses reports usage.output_tokens_details.reasoning_tokens = 0 even when the reply has reasoning #717

What happened

With thinking on, /v1/responses returns a reasoning output item holding the full trace, but usage.output_tokens_details.reasoning_tokens is always 0, both non-streaming and in the streamed terminal events. A client that budgets or reports from this field is told no reasoning happened. When max_output_tokens cuts a reply off, the client cannot tell from usage how much of the budget the reasoning used. /v1/chat/completions has no completion_tokens_details at all, so it gives no count either (#307 and a comment on #302 mention this in passing).

Minimal reproduction

PYTHONPATH=python python -c "
from freetoken.server.generation import GenResult
from freetoken.server.responses_api import build_responses_response, ResponsesRequest
req = ResponsesRequest(model='m', input='hi')
r = GenResult(reasoning='Let me think about this carefully. '*20, content='4', tool_calls=[],
              finish_reason='stop', prompt_tokens=12, completion_tokens=180)
u = build_responses_response(r, req, 'resp_1', 0).usage
print('output_tokens', u.output_tokens, 'reasoning_tokens', u.output_tokens_details.reasoning_tokens)
"

output_tokens 180 reasoning_tokens 0

Cause (current main, 9b68e00)

• python/freetoken/server/responses_api.py:481-490: _usage() hard-codes OutputTokensDetails(reasoning_tokens=0). Every Responses usage object is built by it: the non-stream builder (:452) and the stream's response.incomplete / response.completed / response.failed events (:717, :726, :739, :754).
• Nothing on the generation path counts reasoning tokens. GenDone (python/freetoken/server/generation.py:112) and GenResult (:124-132) carry only prompt_tokens/completion_tokens, and openai_api.py:637-646 _usage() has no completion_tokens_details.

Proposed fix

Count reasoning tokens once in the generation path, carry the count on GenResult/GenDone, and report it as output_tokens_details.reasoning_tokens (Responses) and completion_tokens_details.reasoning_tokens (chat completions). The counter could start in the same state as the reasoning parser (force_reasoning, generation.py:389-422), so that a template that pre-opens <think> is counted from the first token. It would stop at the think-end token. If generation ends before that token, all completion tokens would count as reasoning, so a reply truncated mid-reasoning reports its real count rather than 0. With no reasoning parser, or thinking off, the count stays 0.

Environment

Seen serving nvidia/Qwen3.8-Flash-Next-NVFP4 on an RTX 5080 16 GB; the reproduction above is CPU-only on main 9b68e00.