feat(chat): pace native-agent streaming with an adaptive per-session smoother

[Ukaykhingmarma28]

What

Native-agent (Atlas Agent) responses stream into the chat in visible lumps. A live wire probe against ai.tryatlas.cc confirmed the cause: Claude-via-Vertex emits text in ~400–650 ms bursts (occasional 1–1.5 s stalls), and every hop of our pipeline — gateway, engine SSE decoder, native-agent pump, Tauri emit, RAF batch — faithfully forwards per-delta, so the bursts reach the UI intact. (Full code-read + measurements in research/stream-smoothing.md, repo-local.)

This adds a presentation-only StreamSmoother between the atlas:agents listener and the RAF batcher in App.tsx:

• Per-session char queues drained on a 33 ms tick with an elapsed-time budget, so WebKit-throttled timers (unfocused window) degrade to chunkier-but-current, never a growing backlog.
• Adaptive ratchet-up rate targeting ~450 ms lag (derived from the measured burst gap), with a floor rate so an already-smooth stream passes through at chunk speed.
• Wire order preserved: every non-text delta becomes a queue barrier behind the text ahead of it; turn_finished barriers too, so the tail keeps typing before the idle flip.
• Stop feels like stop: cancelled turns, turn_failed, and agent_disconnected flush synchronously. Permission deltas bypass entirely (modal latency).
• Scope-gated to native-agent sessions; ACP sessions and unknown sessions pass through untouched. The persisted transcript never sees the pacing.

Tests

12 unit tests (stream-smoother.test.ts, fake timers): burst spreading, elapsed-time drain under throttling, floor-rate passthrough, barrier ordering, tail-after-finish, instant cancel/fail flush, per-session independence, surrogate-pair safety.

bun run typecheck clean; the 3 failing suites on this branch (atlas-comms CI-coverage/opt-level guards + a comms midnight test) pre-date this change and are untouched by it.

🤖 Generated with Claude Code