npnormalize::np_addr_freq(), over exactly the rows of a
compiled reference: records per address, per address and suite, distinct
suites per address, suite coverage, distinct names and the top name's share,
and the class. md_publish() ships it in every bundle as
addr_freq.parquet, so a run against a slice of the reference (a state, a
shard) still reads the whole tier's counts, and md_reference() returns it
as addr_freq, the object np_compare(addr_freq = ) takes.