Memorization and incomplete fitting in a recursive arithmetic reasoner rebuilt from its published description: data, predictions and reproducible analysis

Complete arithmetic datasets, per-row saved predictions, evaluation curves, allocation charges, checkpoint identities and deletion records, historical double-precision fixture results, full live-training journals, source code and deterministic table calculations for ten exploratory single-seed recursive and single-pass arithmetic training configurations. Strict accuracy compares canonical operator tuples; semantic accuracy executes any predicted tuple. The repository snapshot and full sanitized scientific records are included. No model weights are distributed. See the included README for multipart reassembly and exact offline reproduction.