Replication package v4 for "When Is Predictive Correction Worth Its Price? A Two-Layer Decomposition of Online Budget Allocation under Corrupted Predictions"

Replication package v4 for the manuscript "When Is Predictive Correction Worth Its Price? A Two-Layer Decomposition of Online Budget Allocation under Corrupted Predictions" (Xuefeng Liu, submitted to Computers and Operations Research). Contains derived estimation panels only — raw competition records are excluded per the organizer's redistribution terms and are available from the organizers on request — together with the full backtest pipeline (30 seeds), two-layer decomposition verification (identity residual 0.0), and machine-checked tables and figures. load_cells() falls back to the derived panel automatically and reproduces every paper number bit-exactly; see README.txt inside for exact reproduction commands. Expected headline: capA p=0.6 correction gain = -0.94% (t29 CI [-5.8%, +6.7%], n=30).