This experimental research study evaluates a leakage-controlled residual-calibrated hybrid model for sequential one-hour-ahead photovoltaic generation forecasting during active station-hours. A source audit of operational records from three solar power plants identified 28,296 raw hourly rows but only 11,736 unique station-hour keys; 16,560 redundant source rows were collapsed only when known plan and actual values were identical, and no conflicting known plan/actual values were found. After restricting supervised modelling to unique station-hours with positive planned generation and known actual output, 4,848 observations remained and were chronologically divided into 3,038 training, 819 calibration, and 991 held-out test observations. The proposed Residual-Calibrated Neural Hybrid (RCNH) combines a two-hidden-layer Tabular MLP with a HistGradientBoosting residual calibrator trained exclusively on the calibration period. On the held-out test period, RCNH achieved MAE = 2,591.80 kWh, RMSE = 3,906.53 kWh, and R² = 0.908609; 95% three-day block-bootstrap intervals were 2,372.75-2,961.81 kWh, 3,537.16-4,505.29 kWh, and 0.848489-0.932518, respectively. Against the train-only Tabular MLP, RCNH reduced MAE by 11.36% (95% block-bootstrap interval 6.98-14.95%; Holm-adjusted two-sided p = 0.003998). However, a matched-label-budget diagnostic favored an MLP refitted on train plus calibration data (MAE = 2,464.63 kWh), and rolling-origin validation showed substantially stronger temporal stability for tree ensembles, with ExtraTrees attaining mean RMSE = 3,064.51 kWh versus 6,106.74 kWh for RCNH. These results show that residual calibration is competitive and beneficial in the final held-out regime but is not uniformly dominant; its effectiveness depends on calibration-window representativeness and temporal regime. The contribution is therefore an experimentally validated, leakage-controlled forecasting protocol and a quantified assessment of where residual calibration helps and where its robustness is limited.