fix: set service.instance.id so metrics carry a unique instance label

SetupOTelSDK now adds service.instance.id (the pod hostname) to
OTEL_RESOURCE_ATTRIBUTES. The OTLP->Prometheus exporter maps it to the
`instance` label on both metrics and target_info, which (1) stops
multi-replica services from colliding on a single series and (2) gives
PromQL joins a unique (job, instance) key. Falls back to the service name
when the hostname is unavailable.

otel.go

@@ -22,7 +22,17 @@ import (
 // SetupOTelSDK bootstraps the OpenTelemetry pipeline.
 func SetupOTelSDK(ctx context.Context, enabled bool, serviceName, buildVersion, environment string) (func(context.Context) error, error) {
 	if os.Getenv("OTEL_RESOURCE_ATTRIBUTES") == "" {
-		if err := os.Setenv("OTEL_RESOURCE_ATTRIBUTES", fmt.Sprintf("service.name=%s,service.version=%s,service.environment=%s", serviceName, buildVersion, environment)); err != nil {
+		// service.instance.id makes every pod a distinct telemetry resource. The
+		// OTLP→Prometheus exporter maps it to the `instance` label on metrics and
+		// target_info, which keeps multi-replica services from colliding on a
+		// single series and gives joins a unique (job, instance) key. Hostname is
+		// the pod name under Kubernetes; fall back to the service name if it is
+		// unavailable so the attribute is always present.
+		instanceID, err := os.Hostname()
+		if err != nil || instanceID == "" {
+			instanceID = serviceName
+		}
+		if err := os.Setenv("OTEL_RESOURCE_ATTRIBUTES", fmt.Sprintf("service.name=%s,service.version=%s,service.environment=%s,service.instance.id=%s", serviceName, buildVersion, environment, instanceID)); err != nil {
 			return func(context.Context) error {
 				return nil
 			}, err