resolveOnDemand deduplicated the cache write but not the upstream
lookup — a burst of concurrent queries for a freshly-learned pool-root
subdomain (A + AAAA from a single getaddrinfo, or multiple peer
workers dialing the same foreign relay) each hit the bypass resolver
independently.
Wrap the lookupRecords call in refreshGroup.Do with an 'ondemand:'-
prefixed key so concurrent callers collapse into one request; the
prefix namespaces this key off scheduleRefresh's keyspace so the two
paths don't collide.
Error and empty-result paths preserved; the cache write still runs
under the existing TOCTOU-safe exists check + mutex so a winning
refresh isn't clobbered.