feat(metrics): single chart engine + full-width stacked layout

- One engine: go-analyze/charts (grafana theme) for all live metrics
- Server chart: CPU temp, CPU load%, mem load%, power W, fan RPMs
- GPU charts: temp, load%, mem%, power W — one card per GPU, added dynamically
- Charts 1400x280px SVG, rendered at width:100% in single-column layout
- Add CPU load (from /proc/stat) and mem load (from /proc/meminfo) to LiveMetricSample
- Add GPU mem utilization to GPUMetricRow (nvidia-smi utilization.memory)
- Document charting architecture in bible-local/architecture/charting.md

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

bible-local/architecture/charting.md

@@ -0,0 +1,38 @@
+# Charting architecture
+
+## Decision: one chart engine for all live metrics
+
+**Engine:** `github.com/go-analyze/charts` (pure Go, no CGO, SVG output)
+**Theme:** `grafana` (dark background, coloured lines)
+
+All live metrics charts in the web UI are server-side SVG images served by Go
+and polled by the browser every 2 seconds via `<img src="...?t=now">`.
+There is no client-side canvas or JS chart library.
+
+### Why go-analyze/charts
+
+- Pure Go, no CGO — builds cleanly inside the live-build container
+- SVG output — crisp at any display resolution, full-width without pixelation
+- Grafana theme matches the dark web UI colour scheme
+- Active fork of the archived wcharczuk/go-chart
+
+### SAT stress-test charts
+
+The `drawGPUChartSVG` function in `platform/gpu_metrics.go` is a separate
+self-contained SVG renderer used **only** for completed SAT run reports
+(HTML export, burn-in summaries). It is not used for live metrics.
+
+### Live metrics chart endpoints
+
+| Path | Content |
+|------|---------|
+| `GET /api/metrics/chart/server.svg` | CPU temp, CPU load %, mem load %, power W, fan RPMs |
+| `GET /api/metrics/chart/gpu/{idx}.svg` | GPU temp °C, load %, mem %, power W |
+
+Charts are 1400 × 280 px SVG. The page renders them at `width: 100%` in a
+single-column layout so they always fill the viewport width.
+
+### Ring buffers
+
+Each metric is stored in a 120-sample ring buffer (2 minutes of history at 1 Hz).
+Buffers are per-server or per-GPU and grow dynamically as new GPUs appear.