### Check for existing issues

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

### What happened?

When a request is sent to the Anthropic-native pass-through endpoint `POST /v1/messages`, LiteLLM forwards the request body to Vertex AI essentially verbatim. Provider/model-specific parameters that the target model does not support are not sanitized — even with `drop_params: true` set globally in `litellm_settings`, and even with no `additional_drop_params` override anywhere in the config.

The proof is that the same model and the same parameter behave differently depending on the route: - `POST /v1/messages` with `model=haiku` (forwarded to `vertex_ai/claude-haiku-4-5`), `thinking:{"type":"adaptive"}` → **HTTP 400** (Vertex rejects it). - `POST /v1/chat/completions` with the same model and same `thinking` payload → **HTTP 200** (LiteLLM seems to drop `thinking` before forwarding; response shows `reasoning_tokens:0`, `thinking_blocks:null`).

This is not because `adaptive` is an invalid value in general — `claude-opus-4-8` accepts `thinking:{"type":"adaptive"}` via the same `/v1/messages` route with HTTP 200. It's a model-capability mismatch (…