## What happens

Calling the proxy's OpenAI-compatible `/v1/embeddings` with `model=gemini/gemini-embedding-2` and an `input` array containing a media data URI (e.g. `data:image/png;base64,…`), optionally mixed with text strings:

1. The response `usage` is all zeros (`prompt_tokens: 0, total_tokens: 0`), even though the upstream native `embedContent` response includes `usageMetadata.promptTokenCount` and per-modality `promptTokenDetails` (e.g. `{"modality":"IMAGE","tokenCount":258}`). 2. The whole `input` array is fused into a single `content` and returns **one** embedding vector, whereas text-only arrays return one vector per input. The single-vector semantics may be intentional for multimodal fusion, but the asymmetry deserves a docs note.

## Why it matters

Google bills gemini-embedding-2 media input at modality-specific token rates (IMAGE 258 tokens/image, AUDIO 25 tokens/s, VIDEO 66 tokens/s — verifiable via `usageMetadata`). With `usage` zeroed, any cost tracking / spend logging / budget enforcement built on the proxy's usage fields silently sees zero cost for media embeddings.

## Expected

Propagate `usageMetadata.promptTokenCount` into `usage.prompt_tokens` / `total_toke…