Admin Data Source Metrics

Counters for how observation endpoints are resolving.

Every request is tagged by which layer served it: local (ingested store), cache (in-process TTL), live (upstream network fetch), or empty (no data anywhere). A healthy platform serves ~100% from local; every live hit is a coverage-gap signal worth triaging (ingestion pipeline hasn't covered that country / indicator / year combo yet).

Also returns a ring buffer of the last ~200 live-fallback events with country + endpoint + latency so ops can see exactly which requests are cold-hitting the wire in real time.