Daemon: opt-in installation, status/recovery UX, benchmarks and acceptance gates

[axisrow]

Part of #14 (delivery step 4). Depends on transport, lifecycle and providers/caches issues.

Что сделать

• Opt-in installation: shared mode writes a client-wrapper command into Claude settings (claude-settings.ts); returning to one-shot restores the previous command. Both directions tested; no TUI/CLI regression (CLI from PR feat(cli): non-interactive subcommands for agent-driven configuration (#602) #11 included).
• Status/recovery UX: daemon status shows protocol/build identity, uptime, aggregate counters, timings — never secrets.
• Platform scope: macOS/Linux first; Windows keeps the existing one-shot path (documented) until an equivalent private transport is tested. Bun/Node compatibility resolved/documented per fork issue Node 14+ distribution requirement is already broken on main (pre-existing blocker) #4 rather than assuming.
• Benchmarks: extend scripts/benchmark-render.py (no new framework). One-shot vs shared at 1/10/36/50 sessions; cold and warm starts; idle timer refreshes; active updates; several repositories/configs/accounts; growing transcripts; slow-command/recovery cases. Report CPU seconds per completed render and per minute including the persistent daemon + all clients/children, aggregate/peak RSS, p50/p95/p99 latency, subprocess counts, provider/cache hits, output equality.
• Acceptance run: execute the epic's acceptance checklist; record honest numbers. The ≥50% aggregate-CPU target in the warmed 36-session workload is a gate to validate, not a claimed result.

Файлы

src/utils/claude-settings.ts, src/ccstatusline.ts, src/daemon/, scripts/benchmark-render.py, docs/USAGE.md.

Эталон

PR #2 / PR #3 (honest benchmark reporting discipline — measured numbers only, no extrapolated claims).

Acceptance (from epic)

Benchmarks report end-to-end cost for daemon and clients/children with unchanged output and equivalent freshness; shared mode remains opt-in with a tested return to one-shot.

Оценка

• LOC: ~400–700 (src ~250 / benchmark+tests ~350)
• Размер: L
• Время (код): ~50–130 мин
• Время (review): ~60–90 мин (база 30–45 ×2: atomic writes to user config + benchmark claims verification)
• Время (итого): ~110–220 мин

Priority: medium.