PCR: gradual reduction in throughput during long initial backfill

A cluster replicating from a 35TB source was observed to gradually be ingesting at slower and slower rates over the first 8 hours of its initial backfill scan. At the 8h mark it was ingesting at an aggregate rate across all 9 nodes of approximately 35MB/s of SST bytes, or about 4MB/s/node, down from about 3x that initially.

Jira issue: CRDB-36642

Start by reviewing the reported 35TB replication scenario and the linked Jira issue CRDB-36642, focusing on the initial backfill scan and its declining SST ingestion rate. No source file or test is identified; progress would require locating the relevant backfill path, reproducing or explaining the throughput decline, and defining a verified improvement.

Written by the indexing model from the issue text.