feat(coding-agent): smooth streamed response pacing

[changeroa]

Summary

• add an 80ms startup jitter buffer for streamed assistant responses
• estimate provider grapheme arrival rate with EWMA and adapt reveal speed between 45–240 graphemes/sec
• retain roughly 140ms of queued text while preserving lifecycle/tool-call flush behavior and 30–120 FPS consistency

Why

The previous fixed catch-up policy emptied provider bursts quickly, producing visible output/pause cycles when the next chunk arrived later. A small adaptive reserve bridges typical chunk gaps without delaying completion.

Verification

• npx vitest run test/streaming-reveal.test.ts (20 passed)
• npx vitest run test/assistant-message-incremental-render.test.ts (11 passed)
• node --test --import tsx --import ./test/setup-multiplexer-env.mjs --test-reporter=dot --test-reporter-destination=stdout test/tui-render.test.ts (45 passed)
• npm run check
• real CLI TUI smoke via senpi-qa (5/5 passed)

QA receipt: local-ignore/qa-evidence/20260727-adaptive-streaming-buffer/

---

Summary by cubic

Smooths streamed assistant response pacing with an adaptive buffer and learned rate control to eliminate bursty output/pause cycles. Keeps ~140ms queued and follows the provider’s arrival rate (no hard cap) for steady, readable updates.

• New Features
  • Adds an 80ms startup buffer to gate the first reveal.
  • Learns arrival rate via EWMA with bounded samples (min 45 u/s); removes the 240 u/s cap and limits extra catch-up to 600 u/s.
  • Maintains ~140ms reserve with a 267ms convergence window; frame-rate independent pacing (30–120 FPS) with fractional carry, and resets carry after fully drained bursts.
  • Splits content and pacing into streaming-reveal-content.ts and streaming-reveal-pacing.ts; exports helpers from streaming-reveal.ts.
  • Shares pacing with tool-arg streaming; parses JSON in bounded batches while preserving surrogate pairs and lifecycle/tool-call flush behavior.

^{Written for commit 766ea19. Summary will update on new commits.}