feat: instance usage monitoring with metrics-server integration

[hippoley]

Summary

Implements real-time instance resource usage monitoring by polling the Kubernetes metrics-server API. The instance_usage table already exists in the DB schema (created by the init migration) but had zero implementation — this PR fills that gap entirely.

What's New

Repository Layer

• InstanceUsageRepository with 5 methods: Create, GetLatestByInstanceID, ListByInstanceID, ListLatestPerInstance, DeleteOlderThan
• ensureTable() as an idempotent safety net (table already exists in migration)
• ListLatestPerInstance() uses a correlated subquery with MAX(recorded_at) for efficient admin overview

Service Layer

• InstanceUsageService with input validation, 24h default / 30-day cap on history queries
• InstanceUsageCollector — background goroutine that periodically collects metrics for all running instances

Collector Design

• Polls K8s metrics-server via clientset.RESTClient().AbsPath("/apis/metrics.k8s.io/v1beta1/...") — no new go.mod dependencies
• Graceful degradation: when metrics-server is unavailable, records uptime only (no crash, no error spam)
• Hardened with sync.Once Stop(), panic recovery, sync.Mutex.TryLock() overlap guard
• Configurable interval via USAGE_COLLECT_INTERVAL env var (default: 60s)
• CPU reported as percentage of instance's allocated cores (millicores → percent)
• Memory reported in GB
• Uptime calculated from pod startTime
• GPU and disk left as nil (requires DCGM / kubelet stats API — planned for Part 2)

API Endpoints

Method	Path	Scope	Description
GET	/instances/:id/usage/current	User	Latest usage snapshot for an instance
GET	/instances/:id/usage/history?hours=24	User	Historical usage records (default 24h, max 30d)
GET	/admin/instances/usage/summary	Admin	Latest usage for all instances

Wiring

• Repository, service, collector, and handler initialized in main.go
• Collector starts after sync service, stops during graceful shutdown
• Routes registered under existing instance and admin groups

Testing

• 11 new tests covering CPU/memory quantity parsing (millicores, nanocores, Mi/Gi/Ki/Ti, SI units), service logic, and edge cases
• All existing tests pass with zero regressions

Design Decisions

1. No new dependencies: Raw REST calls to metrics-server instead of adding k8s.io/metrics
2. Non-breaking: Pure additive — no existing code modified except main.go wiring
3. Graceful degradation: Works without metrics-server (uptime-only mode)
4. Same hardening patterns as backup_scheduler: panic recovery, overlap guard, safe Stop()

Files Changed

File	Change
repository/instance_usage_repository.go	New
services/instance_usage_service.go	New
services/instance_usage_collector.go	New
handlers/instance_usage_handler.go	New
services/instance_usage_service_test.go	New
cmd/server/main.go	Modified (+11 lines)