Selects which statistics the server computes per edge.
A subset of ServiceStatistic. duration.min, duration.max, duration.sum, duration.p95 and duration.p99 are
absent because the connection metric is an explicit-bucket histogram
whose boundaries reach 1200s in a handful of steps; a p99 read off it
is a bucket boundary rather than a measurement.
count: total request count on the edge.
statusCounts: the per-status request breakdown, as a whole. It has
no leaf members, matching ServiceStatistic: the three counts are a
closed partition produced by a single aggregate.
duration: prefix alias for all four duration fields.
The same three rules as ServiceStatistic govern the vocabulary: the
list is a set, so duplicates and overlaps collapse rather than erroring;
no member implies any other, and the server never widens a request to a
superset; and cost is documented, never encoded.
Cost here is steeper than on the node side. count and statusCounts
are one query together. duration.avg is one more. Each named
percentile is one more query on top of that, and each reads the full
bucket set for every connection in the time range. Naming all three
percentiles costs four more queries than naming none.