Node confidence is not normalized by the build (edges and hyperedges are) #4529

A consistency gap in the confidence handling: _fold_confidence normalizes the confidence/confidence_score pair on edges (build.py ~L540) and hyperedges (~L1223), but nodes get no equivalent fold. Node dicts pass through the build verbatim (validate.py checks only {id, label, file_type, source_file}; unknown fields are preserved), so when a prompt schema emits per-node confidence (as ours now does) the values reach graph.json unnormalized: a model emitting "extracted" (lowercase), a bare float, or a drifted string lands as-is, while the same drift on an edge is folded to the enum + score.

Suggested shape: fold nodes in the same place node aliases are folded (_fold_node_aliases), so nodes and edges share one normalization contract. No behavior change for graphs without node confidence.

Environment: graphifyy 0.9.92, Python 3.13. Found wiring node-level confidence through an extraction pipeline; our extractor carries a local backstop fold meanwhile.