A 12.5M-row run is measured in tens of hours, so it cannot be a single call that loses everything on an interrupt. This driver splits the work into batches, writes each batch's result to disk as it completes, and skips finished batches on restart.
What is computed once and shared
The BMF reference — read, normalize, name-frequency table, token IDF, blocking index, and the stage-4 recovery keys — is a function of the reference alone. It is built once and handed to every batch, which is the difference between a 5-minute setup and a 5-minute setup per batch.
Batching by identity, not by row
Batches are formed over deduplicated recipient_uid values from pf_dedup(),
so a cluster is never split across batches and every row of a cluster inherits
the one match computed for it (see pf_propagate_matches()).