v0.9.0 Performance: TUI lag and memory pressure from high agent fan-out sessions

[Hmbown]

Problem

When 30+ sub-agents are running in parallel (as in the v0.8.68 development session on 2026-07-03), the terminal UI becomes increasingly laggy. Observed symptoms:

• Typing latency increases significantly
• Terminal rendering stalls
• Memory pressure on the host machine ("my computer is freaking out")
• Agent completion events flood the parent session, bloating the transcript

This makes high-fan-out orchestration unusable for real work — the tool becomes the bottleneck at exactly the scale where it should shine.

Likely causes

1. Agent completion event flooding — Each agent completion sends a <codewhale:runtime_event> with a full summary into the parent transcript. 30+ completions = massive transcript growth
2. Unbounded context growth — No mechanism to summarize or prune completed agent entries from the parent context
3. Render thrashing — The TUI re-renders the full transcript on every event, and 30 simultaneous completions trigger 30 re-renders
4. Worktree filesystem pressure — 40+ git worktrees each with their own build artifacts consume disk and inode resources

Scope

This issue covers the performance/degradation path:

1. Throttle event ingestion — Batch agent completion events into a single render cycle instead of re-rendering per event
2. Compress agent results in transcript — Store only verdict + key stats in the parent transcript; full agent output goes to artifact files
3. Context budget auto-compaction — Trigger compaction when context pressure exceeds a threshold driven by agent fan-out
4. Worktree lifecycle — Auto-delete worktrees after agent completion (or after configurable TTL)
5. Render debouncing — Coalesce rapid TUI redraws during high-event periods

Non-goals

• Not reducing agent concurrency. The goal is to handle high fan-out, not avoid it.
• Not removing agent event data — just not rendering it all in the live transcript.

Acceptance

• 30+ concurrent agents should not cause visible typing latency
• Parent context should not grow unboundedly from agent completion events
• Terminal rendering remains responsive during agent storm
• Worktree disk usage is bounded (auto-cleanup)
• Tests: benchmark transcript growth with N simulated agent completions; measure render latency

Evidence

Observed live on 2026-07-03 running CodeWhale 0.8.67 with 41 sub-agents. Terminal became laggy, memory pressure was high, several agents cancelled themselves due to resource contention.

[Hmbown]

Claude Workflow screenshot follow-up: the UX target for high fan-out is a compact background-tasks phase ledger, not transcript event spam. #4039 tracks that panel. Performance acceptance here should include rendering 30+ worker rows grouped by phase without typing lag or full-transcript re-render thrash.