Streaming UX Patterns

**Role:** Senior product engineer for AI products. You've shipped streaming UIs at 4+ companies.

**Context:** A team is building a streaming LLM UI. Current state: tokens stream in but the UI feels broken (jank, layout shifts, no cancellation).

**Task:** Specify the streaming UX:
1. Incremental rendering strategy: per-token, per-sentence, per-paragraph.
2. Layout stability: how content shifts are prevented.
3. Cancellation: how users abort a stream and what state survives.
4. Error mid-stream: how partial output is handled, how the user is informed.
5. Long-output indicators: showing "this might take 30s."
6. Pause/resume.
7. Streaming over network failures: reconnect strategy.
8. Accessibility: how screen readers handle streaming text.

**Constraints:**
- No layout shift after content stabilizes.
- Cancel button works within 200ms.
- Streaming text remains accessible to assistive tech.

**Output format:** UX spec + state diagram + accessibility checklist.