On your point, turns out PCA alone doesn't tend to help with Bowmore (or the similar Highland Park). The first component you get is basically medicinal-and-smoky, and these two only score 1 for medicinal. They're also probably a bit too honeyed or floral for cluster 4. But clustering on the components and tweaking the starting values can draw Isle of Jura, Oban, Old Pulteney and occasionally Glen Garioch into the smoky cluster. On the other hand, looking at a biplot of the first two components your could argue for two clusters - the six in OP's original cluster 4 vs. the rest.
The most important thing is that this work led OP to Talisker - result. But I suspect a definitive clustering will require more features - so, sadly, someone needs to taste them all again...
-High dimensional dataset, probably some redundant dimension are being "counting twice". Solution PCA your data first.
-sampling artefacts. Run the k-nearest algorithm lots of times and average
If the data points were lower dimension, the clusters would be able to cover the space better. Bowmore is a smoky whisky (score 3), but it has enough interesting other characteristics that means it does not align with the obvious "Islay" cluster (smoky centre of 2.8). I think it's in the wrong cluster, due to redundant fields having too much power.