High within-dataset accuracy does not establish that a dissolved-gas analysis model will remain valid in an independently sourced transformer population, especially when gas magnitudes, case mix, and label structures differ. This study proposes HIRA-DGA, a Hierarchical Invariant-Representation and Abstention framework, and evaluates it using two public real-data sources: 200 Malaysian development records and 25 Ukrainian cases, of which 19 represented single mechanisms and six were mixed or composite. Five common gases were harmonized. HIRA-DGA separates locally calibrated normal/fault screening from portable fault typing: Stage 0 uses centered log-ratio composition plus total gas magnitude; Stage 1 uses three log gas-ratios for thermal-versus-discharge recognition; Stage 2 uses centered log-ratio composition for discharge severity; and Stage 3 returns HOLD for unsupported label structures. Duplicate-aware 10-by-5 grouped validation, and a one-way, no-target-tuning external challenge were used. Median total gas magnitude differed by 65 to 1,322 times across harmonized classes. With logistic regression, direct four-class absolute-log and ratio models achieved 2/19 and 7/19 external accuracy, whereas HIRA-DGA achieved 18/19 (94.7%; Wilson 95% CI 75.4-99.1%). Stage 2 centered log-ratio features classified 11/11 discharge-severity cases versus 5/11 for its absolute-log ablation. A fixed LightGBM robustness check showed the same architectural direction, improving from 5/19 for direct ratio classification to 13/19 for the hierarchical path. All six mixed cases were assigned HOLD. The contribution is a representation-by-decision technology architecture, not a new classifier: it localizes magnitude-sensitive screening, transfers scale-reduced chemistry and makes abstention an explicit engineering output.