feat(llm): pass reasoning effort through to reasoning models

[Jacky628]

Reasoning models accept a reasoning.effort parameter, but there is currently no way to set it, so every call runs at the provider default.

This matters most for the judge roles added in #316. A reviewer or a tournament judge is exactly the kind of task where more deliberation changes the verdict rather than just the wording — it is a scoring decision, not a generation one.

llm.reasoning_effort (xhigh | high | medium | low | minimal | none) is threaded into both the generator and the reviewer client. Empty by default, so nothing changes unless it is set.

The part worth reviewing

The parameter is gated on the model, because non-reasoning models reject an unknown reasoning field with HTTP 400. Getting this wrong would break every run on a gpt-4o-class model, so:

• _supports_reasoning_effort reuses the existing _NEW_PARAM_MODELS set instead of introducing a second model list. The models that need max_completion_tokens are the same reasoning family, and one list cannot drift out of sync with itself.
• It tolerates a provider prefix, so the check still works for openai/gpt-5.4 through OpenRouter — the existing _NEW_PARAM_MODELS call sites do a bare startswith and would miss those.

One consequence to flag: o1 is not covered, because it is absent from _NEW_PARAM_MODELS. I did not add it, since that set also controls the token parameter and I did not want to change unrelated behaviour in this PR — but if o1 should be there, adding it fixes both call sites at once.

Verified

13 model names classify as expected, including provider-prefixed and non-OpenAI ones:

gpt-5.4, gpt-5, o3, o3-mini, o4-mini            -> True
openai/gpt-5.4, openai/o3-mini                  -> True
gpt-4o, gpt-4.1, claude-sonnet-4-6, deepseek-chat, ""  -> False
anthropic/claude-sonnet-4-6                     -> False

The example config still parses to an empty value, and a configured effort reaches the reviewer client built by reviewer_from_rc_config.

Full suite: 2951 passed, 0 failed, 56 skipped.

Independent of #317 and #318; it touches a different part of the example config, so the three should not conflict.

🤖 Generated with Claude Code