## What happened?

When calling the LiteLLM proxy's `POST /v1/audio/transcriptions` endpoint with OpenAI's `gpt-4o-transcribe-diarize` model and **multiple** speaker references, only **one** reference — the **last** one in the request — is actually forwarded to OpenAI. All earlier `known_speaker_names[]` / `known_speaker_references[]` entries are silently dropped.

OpenAI's diarization model accepts up to **4** known speakers, supplied as **repeated multipart form fields** (the `field[]` array convention):

``` -F 'known_speaker_names[]=alice' -F 'known_speaker_references[]=data:audio/wav;base64,...' -F 'known_speaker_names[]=bob' -F 'known_speaker_references[]=data:audio/wav;base64,...' ```

(See OpenAI Speech-to-text guide → *Speaker diarization*.)

Sending the same request **directly to OpenAI** maps all speakers correctly. Sending it **through the LiteLLM proxy** results in only the **last** reference being honored — the repeated `[]` array fields appear to be collapsed to a single (last) value before being re-sent upstream.

## Definitive minimal reproduction

The cleanest proof that this is an array-collapsing bug (and **not** a model/quality issue): send **two references u…