### What happened?

For provider **pass-through** endpoints with streaming (e.g. `POST /anthropic/v1/messages` with `stream: true`), usage/cost is logged **only when the stream completes successfully**. If the upstream stream raises **mid-flight** — request timeout, connection reset, or a provider 5xx mid-stream — the request logs **no usage at all**, even though the provider has already generated and **billed** the tokens produced before the failure.

This causes a systematic, often large, under-count of tracked spend vs the provider invoice — and it's worst exactly where it hurts most: **long agentic streams** (large/cached context + server tools like web_search) are both the biggest token consumers *and* the most likely to hit the request timeout. So the largest requests contribute **zero** to SpendLogs.

**Root cause** — `litellm/proxy/pass_through_endpoints/streaming_handler.py`, `PassThroughStreamingHandler.chunk_processor()`:

```python async for chunk in response.aiter_bytes(): raw_bytes.append(chunk) ... yield chunk # logging is scheduled ONLY here, after the loop completes: asyncio.create_task(PassThroughStreamingHandler._route_streaming_logging_to_handler(...…