Fleet-wide autoscaling signal for KEDA: total active calls + warm buffer.

Unlike /health/active-calls (a per-worker drain signal), this is the whole fleet's active-call count read from Redis, so a single scrape drives the ScaledObject. buffer (query param, in calls) is the warm headroom: KEDA divides value by the per-pod target K, giving desiredPods = ceil((calls + buffer) / K). All scaling policy — K, buffer, min/max — lives in the ScaledObject; this endpoint is pure signal.

When the fleet count cannot be read (Redis down) this responds 503, NOT 0: a failed scrape makes KEDA's HPA hold the current replica count, while a successful scrape of 0 would instruct it to scale toward minReplicas.