### Check for existing issues

- [x] I have searched the existing issues and checked that my issue is not a duplicate.

### What happened?

Calling a model registered under the azure_ai provider (model group azure_ai/deepseek-v4-pro, backed by a DeepSeek-V4-Pro deployment on Azure AI Foundry) with reasoning_effort set in the request silently drops the parameter before the upstream call - no error, no translation, it simply never appears in the outgoing request body. Enabling litellm_debug confirms this: the logged upstream call to Azure omits reasoning_effort entirely even though it was present in the client request.

The underlying Azure deployment itself fully supports reasoning_effort - calling the same model directly against Azure's endpoint (bypassing litellm) with the identical parameter returns a proper message.reasoning_content field with real chain-of-thought, correctly gated by effort level ("none" disables it, "high" produces substantial reasoning output). So this is a litellm-side gap in the azure_ai provider's parameter mapping, not a deployment limitation.

Setting model_info.supported_openai_params to include reasoning_effort (and supports_reasoning: true) has no eff…