Systematically align training precision across three scenarios — model migration (native→FlagScale), internal iteration (self-regression), and hardware migration (NVIDIA→new hardware). Progressive 6-level elimination from structure to forward/backward.
