On The Hidden Biases of Flow Matching Samplers