Source derivatives preserve the checked native calculation
All 24 native controls pass. Across all 348 bounds, point predictions, coefficient matrices and physical/reference bindings match the sealed predecessor exactly. None of the checked unshared error allowances widens.
The new calculation intersects a source-specific mean-value remainder with the existing physical and training bounds. It accounts for numerical center and row differences explicitly, retains the clock direction and keeps all 51 physical/shared columns. No calibration cap or acceptance tolerance changes.
The study used 78 native evaluations, 60,288 logical bound steps and four single-thread workers, completing in 129.67 seconds. These are generated software controls, with no aircraft observations or protected ESA data.
The next gate is the separately frozen six-source long-history comparison. Useful complete prediction ranges and observational confidence remain open.