AMG constructed from parameter tree uses Unsymmetric criterion with RowSum norm.
Unfortunately, I do not recall what the best default was, but I am pretty sure that for a scalar matrix using RowSum is not the best. For some discretizations (was it Q1 on cuboids?) there will be positive off-diagonals that should be treated as weak connections. With RowSum we will not see the sign.
IMHO I would have used Symmetric with FirstDiagonal and think we should add a parameter tree section for the Coarsening criterion.