Aggregate dependency health for uptime monitoring

/v1/health deliberately touches nothing so it stays green through a control-plane incident — the right answer for a load balancer, the wrong one for a human. This one asks each binding a question: SELECT 1 on D1, a head() on R2 (a miss is a healthy answer), GET /healthz over the orchestrator service binding. It adds today's platform model spend and every open alert the scheduled provider check has written. No vendor is called and no customer identifier is returned.