Improve model response handling

The logic for handling the model response and generating decorations while streaming is brute force. There is certainly a better way to handle this.

The issue names no files, tests, or entry points. First locate the streaming model-response and decoration logic, then clarify the desired behavior and scope with maintainers; done requires an agreed refactor with equivalent or improved streaming output and supporting validation.

Written by the indexing model from the issue text.