Recovering ontology-normalized layer classes from vector-native PDF drawings via graph-based learning

Vector-native PDFs preserve geometry but often lose authoring-time CAD-layer organization. This paper presents a document-preserving vector-to-graph pipeline that reconstructs drawing elements and predicts ontology-normalized layer classes. Raw and canonicalized source-layer strings construct supervision but are excluded from inputs. The benchmark contains 381 drawings and 12.1 million vector elements under a three-level hierarchy of 307 fine-grained classes, 36 intermediate groups, and 5 coarse domains. Local geometric context lifts fine accuracy from 0.342 to 0.438, whereas geometry plus text reaches 0.344. Extended single-run training reaches 0.612 fine and 0.777 coarse accuracy. In matched three-seed runs, graph models reach 0.462–0.487 fine accuracy versus 0.416 for an equally informed MLP, while source-group holdouts are weaker. The evidence supports assisted recovery of ontology-normalized authoring-layer organization, not unrestricted engineering-object recognition or unattended automation.