When using `/v1/responses` with streaming, three code paths access `.response.usage` without checking if the response is a dict. Some upstream providers return dict-format responses, causing `AttributeError` in all three locations.

**Affected files:**

1. `router.py` `_extract_partial_responses_usage` — `return completed.response.usage` crashes when `completed.response` is a dict. Causes HTTP 500 mid-stream.

2. `litellm_logging.py` `_get_assembled_streaming_response` — `isinstance(result.response.usage, ResponseAPIUsage)` crashes for the same reason. Tagged `[Non-Blocking]` but silently drops the log entry — successful requests never appear in SpendLogs.

3. `transformation.py` `translate_responses_chunk_to_openai_stream` — `for item in output_items` crashes when `output_items` is None. Triggers when a client calls `/chat/completions` but upstream only supports `/responses` (LiteLLM bridges the format, then crashes during back-conversion).

**Fix:** Add `isinstance(response, dict)` guards before attribute access, and `or []` for None iterable. PR follows.

**Version:** v1.93.0 (`main-stable`)