Fix PaperBench rubric pruning at subtree boundaries

[sylvesterkaczmarek]

Summary

• assign the existing Subtree task category when pruning an internal rubric node into a leaf
• preserve the original task category when the pruned node is already a leaf
• keep rubric IDs, requirements, weights, and fine-grained categories unchanged

TaskNode requires every leaf to have a task category and every internal node to have no task category. prune_to_depth() currently copies self.task_category when replacing a node with a leaf. For an internal node that value is necessarily None, so pruning at the requested depth raises ValueError instead of producing a pruned tree.

Subtree is already a supported category used for approximate subtree grading, so it is the natural category for a collapsed internal node.

Regression coverage prunes an internal root at depth zero and verifies it becomes a valid Subtree leaf.