What are the UX patterns that make streaming LLM responses feel polished?

Token-by-token streaming solves perceived latency but introduces new UX problems — jarring mid-sentence rewrites, incomplete code blocks, cursor flicker, and users who start reading before the response reverses direction. What are the patterns teams use to smooth streaming output: buffering, sentence-boundary flushing, progressive enhancement, abort-and-restart? Looking for what actually works with real users, not theoretical solutions.