### Summary

The proxy ingests **upstream error-response bodies without any size bound**. When an OpenAI-compatible backend returns a very large error body (we hit a ~114 MB HTTP 400 from a backend's request-validation dump), the proxy reads the whole body, runs regex masking over it, and propagates it through the fallback chain and logs. This **stalls the asyncio event loop for seconds** (health/liveness checks time out, the accept queue saturates) and **inflates RSS** — a single malformed upstream response degraded the whole proxy.

### Environment

- litellm `1.85.0`, proxy mode, single uvicorn worker, heavy streaming chat-completions traffic, OpenAI-compatible self-hosted backends with a fallback cascade.

### Where

On an upstream error, the masked-error path reads the entire response body before doing anything else:

- `litellm/llms/custom_httpx/http_handler.py` — `_raise_masked_sync_error` / `_raise_masked_async_error` read the full body (`_safe_get_response_text` / streaming aread) and then run `mask_sensitive_info(...)` over the whole thing. - `litellm/llms/custom_httpx/llm_http_handler.py` — `_handle_error` uses `error_text = e.response.text` (full body).

With a 100 MB+ …