fix: stop response_format/stop leaking into the ranked-choice judge call

[shrdgn]

What & why

ranked.pick_best's internal classifier call (asked to reply with "ONLY the number" of the best panel answer, max_tokens=8) copies the client's request body and only strips tool-calling fields via strip_tool_fields.

That's inconsistent with synthesize.py's judge call, where response_format/stop are intentionally forwarded because the judge's own generated text is the visible answer (documented in README.md and DESIGN.md: "max_tokens, stop, response_format apply to the judge (visible output)"). In ranked.py, though, the classifier's own text is never returned to the client — only the chosen panel response (responses[index].content) is.

If a client sets response_format: {"type": "json_object"} or a stop sequence on the request, it now leaks into this internal ranking call. That can break _parse_choice (e.g. the classifier tries to emit JSON instead of a bare number) or make the upstream reject the call outright — silently defaulting the ranked pick to index 0 instead of the actual best answer.

Fix: pop response_format and stop from the copied body before sending the ranking call, alongside the existing strip_tool_fields.

How it was tested

$ uv run pytest tests/test_ranked.py -q
19 passed in 0.48s
$ uv run ruff check openfusion/ranked.py tests/test_ranked.py
All checks passed!
$ uv run mypy openfusion/ranked.py
Success: no issues found in 1 source file

• ruff check . passes
• pytest -q passes (no live network)
• New behavior has a test (test_pick_best_strips_response_format_and_stop, asserts the upstream mock never receives either field)
• Docs updated if config / request surface / defaults changed — n/a, no surface change, this restores documented behavior
• No secrets, prompts, or response bodies added to logs or metrics
• Quality/cost claims backed by a reproducible bench/run.py number (if applicable) — n/a, correctness fix

Notes for reviewers

Small, isolated change — two lines in ranked.py plus a regression test. Found via an automated repo-review pass cross-checking ranked.py against synthesize.py's (correct, documented) handling of the same fields.

---

Generated by Claude Code