feat: app-usage Prometheus metrics + Grafana dashboard + Helm scrape

[shimoncohen]

Summary

Adds application-usage metrics to the SvelteKit (adapter-node) server, exposes them for Prometheus, and ships a Grafana dashboard.

• Metrics registry (src/lib/server/metrics.ts): prom-client with default process/nodejs metrics plus three custom metrics — http_requests_total, http_request_duration_seconds (labels method/route/status_code) and demo_views_total (labels client/name). HMR double-registration guard.
• handle hook (src/hooks.server.ts): serves GET /metrics (excluded from its own counter) and times/counts every other request, labeled by the templated route.id (bounded cardinality — never the raw URL).
• Demo view counter: demo_views_total incremented in the demo route load, after the demo is validated, so bogus /demo/*/* paths can't create unbounded label series and failed S3 fetches don't count as views.
• Helm scrape wiring: adds the MapColonies mclabels dependency subchart → renders prometheus.io/scrape|port|path annotations (port 8080, /metrics) and mapcolonies.io/* labels, gated on mclabels.prometheus.enabled.
• Grafana dashboard: grafana/maps-playground-dashboard.json shipped as a sidecar-provisioned ConfigMap (gated on metrics.grafanaDashboard.enabled, configurable discovery label). Panels: request rate by route, p50/p95 latency, error ratio, top demos, process mem/cpu.
• Central vitest-setup.ts guard so node-env test files share the global teardown, plus a test pinning the two dashboard JSON copies in sync.

Test Plan

• npm test — 35 passing (8 files), incl. metrics registry, handle hook, demo-view counter, dashboard-sync.
• npm run check — 0 errors.
• helm lint helm clean; helm template helm renders scrape annotations when enabled, none when disabled; Grafana ConfigMap gated correctly.
• Deploy to a cluster: confirm Prometheus scrapes /metrics and the dashboard auto-loads via the Grafana sidecar.

🤖 Generated with Claude Code

[shimoncohen]

OCP smoke test (infra-dev) — scrape confirmed ✅

Built this branch (60f4ae5) and ran it as a throwaway release in infra-dev, then tore it down. From the live pod:

demo_views_total{client="OpenLayers",name="WMTS"} 5
http_requests_total{method="GET",route="/",status_code="307"} 5
http_requests_total{method="GET",route="/demo/[client]/[name]",status_code="200"} 5
http_request_duration_seconds_count{route="/demo/[client]/[name]",status_code="200"} 5
+ prom-client default metrics

• /metrics serves valid Prometheus exposition; all three custom metrics present and incrementing under traffic.
• route label is the templated route id (/demo/[client]/[name]), not the raw path — cardinality-safe as designed.
• Pod carried the scrape annotations a prometheus.io/scrape scraper keys on (scrape: "true", port: "8080", path: "/metrics").
• Cluster Prometheus itself is RBAC-forbidden to me, so real ingestion wasn't observed — confirmation is endpoint + annotations.

Side finding: this branch also fixes the current dev CrashLoopBackOff

The running playground-dev-playground (old image 0.0.1-4) is in CrashLoopBackOff — getDemoIndex throws Error: property description not allowed on the dev S3 index.json, so GET / 500s and the liveness probe kills it. On this branch's image the same route returned 200 in the smoke test, i.e. the newer code tolerates the description field. Merging + redeploying dev should clear that crashloop as a side effect.