One entity batch’s base rows: every row introduces a brand-new master.
Masters are contiguous — row j’s master is global index
first_master_idx + j of entity entity_idx, and its record is global
record index first_record_idx + j (the same numbers the pipeline
formats into master_id/record_id).
View into a contiguous range of a [Bitset], returned per-cluster by
ClusterCsr::groups — parallel to the &[u64] record-index slice for
the same cluster.
Compressed-sparse-row representation of the duplicate-cluster
membership. Built once, in GtAccumulator::finish, from every
(master, record, is_identical) triple belonging to a duplicated master.
The duplicate copies generated from one BaseRows batch. Every copy’s
master belongs to that same batch (pipeline.rs samples a batch’s
duplicates exclusively from its own masters), which is what lets the
whole batch be classified as soon as both are known.