feat: streaming responses

Roadmap priority: B13

Summary

Stream partial responses as they generate, rather than waiting for the complete reply.

Why it matters

Long generations currently block with no feedback — the NetLogo UI freezes for the full duration of the call. Streaming gives progressive output, which matters for demos with visible agent "speech" and for any single call producing a long passage.

Currently all providers explicitly disable it (e.g. OllamaProvider.scala:50 sends "stream": false).

What needs to be done

1. Decide whether this is even a good fit for NetLogo's execution model. A reporter must return a value; streaming implies incremental delivery, which has no natural NetLogo idiom.
2. Plausible shapes: a callback command run per chunk, a mutable buffer the modeler polls, or streaming only into a display widget while the reporter still returns the final string.
3. SSE parsing per provider — all four use different chunk formats.
4. Interaction with historyLock — partial content must not be committed to history until complete.
5. Interaction with thinking models, where reasoning and answer stream as distinct block types.

Open questions

• Is there a real modeling use case, or is this mainly UI polish? Worth deciding before building, since the NetLogo-side design is the hard part rather than the HTTP side.
• If the only benefit is a responsive UI during long calls, a progress indicator may deliver most of the value for far less work.

Start with the NetLogo execution-model question in the issue and inspect OllamaProvider.scala:50, where streaming is disabled. Compare the four providers' response formats and consider the listed historyLock and thinking-model constraints. Done means agreeing on whether streaming fits NetLogo and selecting a concrete delivery shape before implementation.