A colleague wants to replace the pointwise cross-encode...

The deciding argument is not accuracy, it is what the score is a function of. Pointwise scores depend on (q, d) alone, so they cache across queries that share a document, batch across shards, and stay valid when the index gains a document. A listwise score depends on the entire window, so the cache key must include the candidate list - hit rate collapses - and every incremental index update silently invalidates the scores of documents nobody touched. The resolution is a boundary, not a preference: pointwise from 100 to 20 on the hot path, listwise from 20 to 10 only for the query segments that justify the latency.