ADR-0009: OpenTelemetry & Prometheus observability

Status

Accepted

Context

Enterprise operators need mesh health, task throughput, and latency visibility without drowning nodes in telemetry.

Decision

• Prometheus text exposition at /metrics (always with mgmt feature).
• OpenTelemetry OTLP export behind --features otel + --otlp-endpoint.
• Default trace sample ratio 5% (ParentBased + TraceIdRatioBased) targeting <2% CPU overhead at typical load.
• Spans on critical paths: task.execute, HTTP management, transport classes.

Consequences

Default builds stay lean. Full OTEL stack is opt-in.