## What happened?

When `/v1/chat/completions` is bridged to the Responses API (e.g. `azure/responses/gpt-5.5`), hitting `max_output_tokens` causes Azure/OpenAI to emit `response.incomplete` with the real usage (`output_tokens`, `reasoning_tokens`, etc.).

LiteLLM's chat stream bridge only maps usage from `response.completed`. `response.incomplete` (and `response.failed`) fall through as unhandled events, so upstream usage is dropped. `stream_chunk_builder` then falls back to local `token_counter` on the visible text, which:

- under-reports `completion_tokens` (e.g. 607 vs Azure `output_tokens` 1024) - loses `reasoning_tokens` (shows 0) - also mis-counts `prompt_tokens` for multimodal inputs

## Expected behavior

Terminal Responses events that carry usage (`response.completed`, `response.incomplete`, `response.failed`) should translate that usage into the chat stream chunk. Incomplete should map to `finish_reason=length`.

## Reproduction

1. Configure a model as `azure/responses/<deployment>` (or any chat path with `mode: responses`). 2. Call `/v1/chat/completions` with streaming and a low `max_tokens` so the model hits the cap. 3. Observe LiteLLM debug logs: `type=response.inco…