### What happened

LiteLLM proxy `v1.90.2` (`ghcr.io/berriai/litellm-database:main-stable`), a `model_list` entry with `model: openai/<azure-openai-compatible deployment>`, `stream: true`, `drop_params: true`, called through the **Anthropic `/v1/messages`** surface by Claude Code (`claude-cli/2.1.x`). Provider-based routing sends these through the **OpenAI Responses API** bridge (`_should_route_to_responses_api`).

Under burst load (8 parallel agentic subagents + a heavy interactive session) the Azure upstream episodically fails mid-stream. Because the HTTP `200` and `message_start` are committed before the upstream resolves, the failure is converted into a **well-formed, zero-content SSE stream**: `message_start` → `message_delta (stop_reason: end_turn)` → `message_stop`, sub-second, with **no `error` event**. The client cannot distinguish this from a model that legitimately produced no output — Claude Code subagents return a silent empty result (`""`).

This is the same class as #20347 (closed stale) and #24004, but reproduced specifically on the **openai-provider → Responses-API → `/v1/messages`** path, with two additional findings below.

### It is completely invisible server-s…