Public per-node subtree aggregate over all observations routed through this node (leaves: the arm snapshot; splits: the exact Chan-merge of both children plus any RegressionSplitNode.carryover). Lets callers score internal nodes — e.g. a Thompson walk-down that picks a branch by its subtree's posterior — without the tree having to keep a live arm on every split node. Equivalent to a directly-accumulated internal arm: Chan's parallel merge is exact for weighted-variance aggregates.