TRM checkpoints trained with progress–stability co-design for TAPS adaptive recurrent inference. Includes Sudoku and Maze models, config