The changed set is the day's work queue, not the population at risk. Counting new lost records against it made 48 host-blocked repositories look like 16% of the corpus when they are 0.25% of 19,000.
Use the full repository count as the denominator for both incremental and full runs. Keep the numerator at new publications/documents with no previous records, and keep the existing 10% threshold, so a genuinely broken run still fails.
180 tests and prek pass.