Streaming-tool fixes must preserve progressive text in tool-enabled requests

Streaming-tool fixes must preserve progressive text in tool-enabled requests. A server patch that drains generation and replays canonical function-call events after completion can make SSE shape tests pass while silently destroying text TTFT for voice agents. Validate both (a) structured function-call events with realistic tool lists and (b) a normal-text response with the same tools, checking that the first text delta precedes completion in real time. This applies to non-Harmony Responses serving and similar adapters; it does not imply every model will produce text under auto tool choice.