Doing some work on aromaticity algorithms and seeing if I can replace the atom type based CDK model with a simpler (and faster) equivalent. I got down to only 122 difference in ChEMBL and on closer inspection it looks to be due to wobbly atom types.

Simple example:

C1=CS[S+]([O-])S1
C.sp2 C.sp2 S.3 S.inyl.charged S.3 (calling percieveAtomTypesAndConfigureAtoms once)
C.sp2 C.sp2 S.planar S.inyl.charged S.planar (calling percieveAtomTypesAndConfigureAtoms twice)

The issue is some of the checks rely on properties that are order dependant. In this case S.3 becomes S.planar if it is next to an Sp2 atom.... it checks for a double bond or Sp2 hybridised. However this isn't quite sufficient as we can see.

Not quite sure what the fix is.