fix: omit temperature and use max_completion_tokens for o-series models

[clifton]

Problem

The OpenAI o-series reasoning models (O4Mini, O3, O3Mini, O1, O1Pro) could not actually be called. Every request through the shared OpenAI-compatible request struct unconditionally serialized temperature and only ever used max_tokens — but on /v1/chat/completions the o-series models return 400 for temperature and require max_completion_tokens instead of max_tokens. There was even a unit test asserting "temperature must always be present", locking in the broken contract.

Fix

• OpenAIModel::is_reasoning_model() — new capability predicate matching all o-series variants. Custom identifiers are matched by prefix (o1-/o3-/o4-), so models like o3-pro without a named variant are detected too.
• Shared request struct (src/backend/openai_compatible.rs): temperature is now Option<f32> with skip_serializing_if, and a new optional max_completion_tokens: Option<u32> field was added.
• OpenAIClient::request_tuning() — centralizes per-model parameter shaping, used by all three OpenAI request builders (materialize, generate, streaming):
  • o-series: temperature omitted entirely; the configured max_tokens is sent as max_completion_tokens; reasoning_effort is sent from the configured thinking level (Off is omitted rather than sent as "none", which o-series models reject).
  • GPT-5.x: unchanged — reasoning_effort from the thinking level, temperature forced to 1.0 while reasoning is enabled, max_tokens as before.
  • All other models: unchanged — temperature and max_tokens sent exactly as configured.
• Grok shares the request struct but has no o-series models: its requests still always send temperature and max_tokens, and never max_completion_tokens.
• The tool-calling loop (tools feature) builds its own JSON body in tools.rs and is untouched (out of scope).

Tests (all offline — no live API calls)

• Updated the "temperature must always be present" serialization test to the new contract; added unit tests for omit/include of temperature and max_completion_tokens.
• New mockito request-body tests in tests/http_mock_tests.rs:
  • o3 (materialize): no temperature, no max_tokens, max_completion_tokens: 1234, reasoning_effort: "medium" (default thinking level).
  • o4-mini with ThinkingLevel::Off: no reasoning_effort/temperature/max_tokens/max_completion_tokens.
  • gpt-4o: temperature and max_tokens unchanged, no max_completion_tokens.
  • Grok: request body unchanged (temperature + max_tokens always present).

Verification

• cargo fmt — clean
• cargo clippy --all-targets (default features, matching CI) — clean; also clean with --all-features
• cargo test — full suite passes (146 lib unit tests, all integration tests, 63 doctests)

🤖 Generated with Claude Code

[clifton]

Closing in favor of removing the o-series models from the Model enum entirely — they're deprecated, and special-casing temperature/max_completion_tokens for them isn't worth carrying. Replacement PR incoming. Custom model IDs remain available for anyone who still needs to call these models.