Preserve group scoring when partitioning passage search

When replacing exhaustive passage scoring with partitioned approximate search, preserve the existing group aggregation before candidate truncation. Taking a small top-k of raw passages inside each partition can allow one long source to occupy the entire budget and hide other useful groups. Aggregate passage scores by group within each lane, merge group maxima across partitions, then select lane candidates and apply rank fusion. Once candidate groups are chosen, rescan all their passages for evidence. This preserves aggregation semantics over the scanned set; it cannot recover groups excluded by approximate routing. Hypothesis: this layout can reduce score-buffer traffic on tensor hardware, but speed and supported-passage recall require matched measurement.

Applies when: A passage index returns source groups and is being partitioned for accelerator search. Keywords: tensor vector search partition passage session grouping reduction recall