fix: send max_completion_tokens to GPT-5 and o-series models - #155

Summary

Fixes #153. Sampler calls go out through OpenAICompatibleProvider.chat() / chat_stream(), which copied the caller's kwargs straight into the request. With an OpenAI reasoning model (gpt-5, gpt-5-mini, gpt-5.4, o3, …) every call was rejected with a 400 — Unsupported parameter: 'max_tokens' is not supported with this model. Use 'max_completion_tokens' instead — so no algorithm was generated, and the provider connectivity test in the web UI failed the same way with a valid key.

Changes

In src/llm4ad/infra/provider/openai_compatible.py:

• New _normalize_request_kwargs(), applied in both chat() and chat_stream() before the request is built.
• For GPT-5-family and o-series models (the gpt-5-chat variants excluded — they are plain chat models), a caller's max_tokens is renamed to max_completion_tokens; an explicitly passed max_completion_tokens wins.
• For the o-series and the original GPT-5 generation (gpt-5, gpt-5-mini, gpt-5-nano, including dated snapshots), which reject any temperature other than 1, a non-default temperature is dropped so the request does not fail on the next parameter. Later GPT-5.x models keep their temperature.
• Every other model keeps the caller's request shape exactly as before.

Tests

Eleven new request-shape tests in tests/unit/provider/test_openai_compatible.py, using that file's mocked-client style:

• gpt-5.4-mini chat: budget arrives as max_completion_tokens=4096, no max_tokens, temperature kept;
• o3 / o4-mini / gpt-5 / gpt-5-mini / gpt-5-nano: budget renamed and non-default temperature dropped;
• gpt-5 with temperature=1: kept (it is the models' own default);
• guards: gpt-4o, gpt-5-chat-latest, and deepseek-chat pass max_tokens and temperature through unchanged;
• chat_stream with gpt-5.4-mini: same normalization on the streaming path.

Verification: the provider test file goes from 24 to 35 passing; with the provider change reverted, 8 of the new tests fail (the 3 guards stay green by design). A full pytest run shows a failure set identical to the unmodified baseline (9 pre-existing failures in unrelated AST/evaluator/settings tests), with exactly the 11 new tests added to the pass count. ruff check and mypy are clean on the changed files.