AmgSymMatrix, despite the name, does not actually fill symmetrically: you have to go through and fill (i, j) and (j, i) separately. Someone (almost certainly me) overlooked this, probably during CLHEP->Eigen migration.
As a postscript, I did supervise a project a few years ago to get Eigen matrices filled symmetrically and stored in half-diagonal form, but it did not fully converge: https://indico.cern.ch/event/738804/contributions/3106176/attachments/1700771/2738969/GSoC2018_ACTS_Developer_Meeting.pdf
Stewart Martin-Haugh authored