token streaming

Iin `server.ts` we could broadcast Ai tokens to the room before saving the document. Doing it smart would require reserving the document id up front but not saving it for each token.

Start by reading server.ts and trace how AI tokens are generated, broadcast to the room, and saved as a document. Investigate how document IDs are currently assigned and how room updates are delivered. Done means tokens reach the room before the completed document is saved without creating a document for every token; add or update focused tests if the existing entry points support them.

Written by the indexing model from the issue text.