The board EM solver writes lossless S-parameters on purpose, and nothing puts the loss back, so every Touchstone file overstates S21 and every downstream check (passivity, eye diagram, measured vs predicted) inherits it. On 100 mm of FR4 at 6 GHz that is about 1 to 2 dB missing, and the stackup parser also ignores KiCad's loss_tangent while the stackup table keeps tan d only as text.