D-29: Dialogue — performance profiling + LRU cache tuning

[kirich1409]

Description

Profile Dialogue under realistic load — large transcripts, rapid token streams, long tool outputs — and tune throttling / caching / virtualization to keep UI responsive.

Spec: Epic #250 §11 Risks rows 6-7.

Scope

Scenarios

1. Rapid token stream — mock session producing 100 text_deltas/sec for 30s. Expected: UI stays responsive, FPS ≥ 30, main-thread CPU < 50%.
2. Long transcript — 500+ messages accumulated. Scroll performance should stay smooth (LazyVStack + .id() stability).
3. Large tool output — Bash returning 1 MiB stdout. Expected: truncation works, no UI freeze; expand to full result opens quickly.
4. Many tool calls — 50+ tool cards in one message. Expected: card rendering stays under a frame budget.
5. Markdown re-render — 50KB single assistant message with mixed content; streaming mode.

Instrumentation

• Use Instruments (Time Profiler, SwiftUI instrument, Memory).
• Add signpost markers around:
  • Parser ingest per event.
  • Reducer action processing.
  • Markdown parse + render.
  • Scroll updates.

Tuning

• MarkdownBlockCache LRU size — default 128, verify no eviction of hot blocks under normal load.
• Throttle frequency — 30fps default; verify perceived smoothness.
• IdentifiedArrayOf<AgentMessage> update cost — ensure no full-array copy on textDelta.
• Scroll virtualization — LazyVStack stability via stable .id() per message.

Escalation

If metrics fail significantly — escalate to performance-expert agent with profiling artifacts.

Acceptance Criteria

• All 5 scenarios profiled; results documented in swarm-report (or inline in PR description).
• Meets targets:
  • FPS ≥ 30 during streaming.
  • Main-thread CPU < 50% in scenario 1.
  • Memory growth < 100MB in scenario 2 (500 messages).
  • No visible jank when expanding / collapsing tool cards.
• Cache hit rate ≥ 80% in scenario 5.
• Any regressions after tuning gated via tests (e.g., throttle frequency assertion).
• Follow-up issues filed for non-blocking perf items.

Relationships

• Parent issue: EPIC: Dialogue — structured chat UI for agent sessions #250
• Blocked by: #D-13, #D-18, #D-28 (test infrastructure to reproduce scenarios)
• Blocks: release (MVP acceptance)