feat(perf): always-visible main-process heap + rss badge in header

[Juliusolsson05]

Summary

Adds a 1 Hz live readout of the Electron main-process v8 heap and RSS to the header strip, gated on `CC_SHELL_PERF=1`. Renders null when the flag is off, so default users see no change.

Motivated by today's OOM that crashed at ~2.69 GiB / 2.78 GiB cap with two back-to-back Mark-Compact GCs freeing 0 MB each — i.e. fully-retained heap, not a classic leak. With this badge sitting in red 30+ minutes earlier, you'd have had a visible warning instead of a sudden kill. PR #48's heap-watchdog already dumps a forensic snapshot at the same 80% threshold; this PR is the user-facing counterpart so growth is visible BEFORE the snapshot fires.

What you see

Badge in header (always visible when CC_SHELL_PERF=1):
```
heap 2.10 GB / 4.00 GB ▁▂▃▄▅▆▇ rss 3.2 GB
```

• Heap color: green <60% of cap, yellow 60–80%, red >80%
• Inline 60-second sparkline of heap-used ratio (same color as text)
• RSS in muted text (no cap to color-zone against)

Click expands to popover (same 430px / border / shadow / position as the existing PerformancePanel):

• Current heap / RSS / external / arrayBuffers
• Peak heap + RSS over the buffered 10-minute window
• 60s growth rate for each metric (positive growth highlighted)
• Overlaid SVG line chart: heap-used (green) + RSS (sky), with a dashed red heap-cap reference line

Implementation

• `shared/performance/types.ts` — new `SystemPerformanceStats` type with WHY comment explaining why it's NOT `PanePerformanceSnapshot` (per-agent) or `PerformanceSnapshot` (on-disk logs).
• `main/ipc/performance.ts` — new `performance:system-stats` handler returning `process.memoryUsage() + v8.getHeapStatistics()`. Gated on `performanceService.getConfig().enabled`.
• `preload/api/performance.ts` — new `getSystemPerformanceStats` method.
• `renderer/src/features/system-perf/` (new dir):
  • `useSystemPerfPoller.ts` — probe-then-poll. First tick decides whether to schedule the 1 Hz interval; if the flag is off no interval is started. Ring buffer of 600 samples (10 min).
  • `ui/SystemPerfBadge.tsx` — header strip + sparkline.
  • `ui/SystemPerfPopover.tsx` — expanded view, mirrors `PerformancePanel` chrome (430px / right-anchored / shadow-xl).
  • `ui/SystemPerfHeader.tsx` — orchestrator. Owns local open state.
• `app/App.tsx` — mounts `` next to the existing perf toggle in the title bar.

What this is NOT

• Not per-agent telemetry. `PerformancePanel` already covers per-pane cpu/memory.
• Not tunable. Zone thresholds (60/80%) hardcoded for now.
• Not persisted. Ring buffer is renderer-only, in-memory.
• Not in the command palette. Click on the badge is the only entry point.

Test plan

• `npm run build:app` clean (13.3s, no new TS errors)
• Manual: `CC_SHELL_PERF=1 npm run dev`, observe badge populates within 1s
• Manual: idle / load many sessions, watch sparkline grow
• Manual: click badge, popover opens with growing chart
• Manual: `CC_SHELL_PERF=0 npm run dev`, confirm badge does NOT appear
• Manual: trigger high heap pressure, confirm zone transitions green → yellow → red

🤖 Generated with Claude Code