Without looking at it, my guess is that the variant changes the original stop codon to a different stop codon (that is, it switches among TAG, TAA, and TGA). So the codon changes (it's non-synonymous), but the effect for the gene should not.

Edit: What I described is a SYNONYMOUS_STOP. As I understand it, a NON_SYNONYMOUS_STOP can't arise, and exists in the code only as an oversight. As evidence of this, only a SYNONYMOUS STOP is explicitly documented in the FAQ. Furthermore, in one of my datasets, I observe over 1000 SYNONYMOUS_STOP calls and no NON_SYNONYMOUS_STOP calls.

More explanation from the relevant code:

                        if (aaOld.equals(aaNew)) {
                                // Same AA: Synonymous coding                                                                                                                                                                                 
                                if ((codonNum == 0) && codonTable.isStartFirst(codonsOld)) {
                                        // details removed
                                } else if (codonTable.isStop(codonsOld)) {
                                        if (codonTable.isStop(codonsNew)) effectType = EffectType.SYNONYMOUS_STOP;
                                        else effectType = EffectType.STOP_LOST;
                                } else effectType = EffectType.SYNONYMOUS_CODING;
                        } else {
                                // Different AA: Non-synonymous coding                                                                                                                                                                        
                                if ((codonNum == 0) && codonTable.isStartFirst(codonsOld)) {
                                        // details removed
                                } else if (codonTable.isStop(codonsOld)) {
                                        if (codonTable.isStop(codonsNew)) effectType = EffectType.NON_SYNONYMOUS_STOP;
                                        else effectType = EffectType.STOP_LOST;
                                } else if (codonTable.isStop(codonsNew)) effectType = EffectType.STOP_GAINED;
                                else effectType = EffectType.NON_SYNONYMOUS_CODING;
                        }

There is no way for the old and new codons to be different (to get into the outer else block) and for both codons to be stop. If they're both stop codons, they'll be the same and go to the first if branch. So the NON_SYNONYMOUS_STOP branch is never reached.