[moetemp]

What was changed

This PR adds a small chat app in tests/contrib/streams/examples/chat_app/ that uses most parts of the Streams API.

• Each chat is a ChatSession Workflow with an agent loop. One message can take several rounds: a model round, tool calls, tool results, the next round, and the answer. Each round streams its tokens to a tokens Stream, and each round, tool call and tool result is a Turn in the turns Stream.
• Tools run as Activities and write their own progress and result Turns. Two tools in one round write to the same Stream at once, so the log shows two Producers.
• An add_message Signal adds a message during a run. It shows the Turn at once with publish_nowait, and the message joins the next round.
• A research tool starts a child Workflow that creates its own Streams under ids it picks, and passes their refs to the parent. The page opens it as a nested panel.
• A fail toggle fails a round's first attempt, so the retry's Superseded item crosses out the failed tokens. A reload or a server restart resumes each read at the tab's saved offset. Continue-as-new happens between turns, and the next run keeps the Streams.
• "Share a snapshot" copies the chat into a new closed Stream with a strict client.create_stream. Each copy keeps the Turn's offset and Producer in the chat.
• A scenario bar has one preset per case, and "Run all" plays them in one chat.
• The model is Claude (with tool use), OpenAI or a scripted model that plans the same rounds, by CHAT_MODEL_PROVIDER or by which key is set. The app reads no key itself, and a failed call keeps only the error's class.
• The app's README.md lists where each API feature is used.

Why?

It shows how the API reads in a real app, not only in examples. The app also found a bug in the Workflow _nowait writes, which #377 fixes.

The app uses today's names (publish, subscribe, BufferedPublisher, create_or_get_stream). It changes with the renames after the API review. A live Claude run with tool rounds works. OpenAI was checked only against a local stand-in. The fix for an end signal during continue-as-new has no test, since the race can't be timed reliably.

Part of the Streams review series.

Checklist

1. How was this tested:
   • poe lint, and the Streams suite with the app tests on both backends with Redis.
   • Each scenario in headless Chrome with the scripted model, and a live Claude run with tool rounds.

[moetemp]

Closed in favor of #381, the combined branch, which has this change.