What Went Wrong? Closing the Sim-to-Real Gap via Differentiable Causal Discovery