Export CPU utilization in observe.Snapshot

[FumingPower3925]

Problem

The overload middleware needs CPU utilization to make shedding decisions. CPU monitoring exists in internal/cpumon (Linux: /proc/stat, other: runtime/metrics) but is not exposed through the public API.

Unblocks

• Overload Management middleware (fix: io_uring send queue corruption under concurrent writes #34) — 5-stage CPU degradation ladder

Change

// observe/collector.go — add to Snapshot:
CPUUtilization float64 // 0.0–1.0; -1 if unavailable

// observe/collector.go — add CPU monitor support to Collector:
func (c *Collector) SetCPUMonitor(m CPUMonitor) { c.cpuMon = m }

// observe/cpumon.go — public interface + constructor:
type CPUMonitor interface {
    Sample() (float64, error)  // returns 0.0–1.0
}

// NewCPUMonitor() CPUMonitor — returns platform-appropriate implementation
// (linux: /proc/stat, other: runtime/metrics)
// Thin wrapper around internal/cpumon to keep the internal package internal.

In Snapshot(), if cpuMon != nil, call Sample() and populate CPUUtilization. Otherwise set to -1.

Performance

CPU sampling only happens during Snapshot() calls (Prometheus scrape, ~every 15s). Zero cost on the hot path. The cpuMon field is a nil-checked interface on Collector, not on Context.

Files

• observe/collector.go
• observe/cpumon.go (new)
• internal/cpumon/ (keep internal, wrap)