Both ternary branches in planShardAssignments() omitted weight from the
pushed entry objects despite using entry.weight for bin-packing. Downstream
consumers that relied on shard entry weight (e.g., balance verification)
had to reconstruct it from computeSplitPlan output.