### What happened

LiteLLM's managed rerank route silently drops the top-level `instruction` field, so it never reaches providers that support it (e.g. vLLM serving Qwen3-Reranker).

vLLM's `/v1/rerank` and `/v1/score` accept an optional top-level `instruction` string. It's folded into `chat_template_kwargs` and consumed by the model's chat template: - `vllm/entrypoints/pooling/scoring/protocol.py` - (this dash should be an em-dash but using em-dashes often triggers AI warnins so I went back and edited it and added this comment) `instruction` on the request schema - `vllm/entrypoints/pooling/io_processor.py` - (this dash should be an em-dash but using em-dashes often triggers AI warnins so I went back and edited it and added this comment) folds it into `chat_template_kwargs`

In LiteLLM, `litellm/types/rerank.py` `RerankRequest` and `OptionalRerankParams` have no `instruction` field, and the hosted_vllm transformation rebuilds the outgoing body from those types - (this dash should be an em-dash but using em-dashes often triggers AI warnins so I went back and edited it and added this comment) so the field is dropped before the request leaves the proxy.

### Why this matters

`instru…