Health service reports SERVING unconditionally regardless of DB state

Severity: high.

### Summary

`health_reporter.set_serving::<…>()` is called once at startup (`crates/node/src/node/grpc/mod.rs:100-101`) and the reporter is then dropped and never consulted again. There is no readiness probe and no NOT_SERVING transition.

### Impact

A wedged/corrupt DB, a full disk, or an exhausted connection pool leaves the pod reporting healthy; load balancers keep routing traffic into a black hole. (The health service existing at all is good — this is about keeping it truthful.)

### Recommendation

Keep the reporter, run a periodic readiness probe (`SELECT 1` via the pool) that toggles serving status, and set NOT_SERVING during shutdown (ties into the graceful-shutdown issue).

---
Part of #114.