fix(llm): temperature is always sent and rejected by claude-opus-5-5 (breaks the #700 federation path) #701

Problem

cora always sends temperature in the chat-completions request (src/engine/llm/transport.rs:270, "temperature": config.temperature, default 0.0, an f32 that cannot be unset). Newer Anthropic models reject the parameter, so every review against them fails:

LLM API returned status 400: {"error":{"code":"invalid_request_error","message":"`temperature` is deprecated for this model.","type":"invalid_request_error","param":null}}

This makes the workload identity federation path from #700 unusable with its default model (claude-opus-5-5): the token exchange succeeds, then the review fails. The Cora Review check still turns green, but the verdict is "Review could not complete", which blocks a merge gate that requires a real verdict.

Reproduction

codecoradev/tole#449 ports #700 to tole (pull_request trigger, so the PR's own run uses the new workflow):

• Run 40462794838 (head fedb691): step "Exchange the GitHub OIDC token…" logs Using a short-lived Anthropic access token (workload identity federation).
• "Run Cora AI Code Review" with cora-base-url: https://api.anthropic.com/v1, cora-model: claude-opus-5-5 (cora 0.19.0): Attempt 1: exit=0, sarif_bytes=612 (fallback SARIF), retries, and the PR comment ends as the 400 above.

The only difference from the passing #700 smoke test (anthropic-wif-smoke.yml, HTTP 200 on /v1/chat/completions with the same token type and model) is the temperature key, which the smoke request does not send. Expected minimal reproduction (not run outside CI, since it needs a federated token):

curl -sS https://api.anthropic.com/v1/chat/completions \
  -H "Authorization: Bearer $TOKEN" -H "Content-Type: application/json" \
  -d '{"model":"claude-opus-5-5","max_tokens":16,"temperature":0,"messages":[{"role":"user","content":"ok"}]}'

Proposal

• Make temperature optional: Option<f32> in LlmConfig/the engine config, sent only when set (in .cora.yaml llm.temperature or the global config), and keep it in the cache key as None vs Some.
• Or, on a 400 that names temperature, retry once without it (and log that it was dropped).
• Consider a model capability check so cora config validate can warn about a temperature the model will reject.

Acceptance

• A review with base_url: https://api.anthropic.com/v1 and model: claude-opus-5-5 completes with a real verdict.
• Providers that accept temperature keep deterministic 0.0 output when it is configured.
• A regression test asserts the request body has no temperature key when it is unset.