Cactvs should be 100% clean for any structure limited to E/Z bond stereochemistry (inc FE participation), tetrahedral (inc. FE participation) atom-centered stereochemistry, even and odd allenes (including analogs with rings instead of double bonds, as long as the stereo center is still located on an atom or bond, and not in free space), and square planar stereochemistry. This includes trans double bonds in large rings (and as far as I know, this is the only software which can handle this), and sequences of cis/trans bonds with non-sequential atoms (difficult for software because /\ SMILES bond stereo markers need to be shared for stereo locations). The toolkit verification procedures include a back-and-forth transformation run with 1000 multi-stereocenter natural products between various formats.
Counterexamples can probably still be found and reports are welcome ;–)
Note: A problem in the Cactvs SMILES encoder for cpds where a stereo atom with an FE which is simultaneously the bridgehead of a cage structure and one or more ring links follow directly after the stereo bridgehead atom in the string and the ring link atom(s) was/were encoded in the SMILES string before the bridgehead atom, sometimes could yield a wrong stereochemistry in the SMILES output. This is fixed in version 3.384 and later.
Anybody endeavoring on the comparison of SDF vs. SMILES representations should read two specification parts very carefully:
a) The rules for stereochemistry in SMILES wrt ring links
b) The parity rules in SDF files wrt hydrogens which are not the last in the list of stereo center neighbor ligands
I know of plenty of softwares which do not get these right.