A label that was not in the fit
A mixture or a clustering fit uses the four iris measurements and does not see the species label. The label is compared with the fitted groups afterwards.
Why is agreement with that label not a prediction for a new flower?
Worked answer
The label was held out, so the agreement describes these 150 flowers. It is not a rule trained to predict species. The flowers are not a random sample of every iris, and a component that lines up with a species in this table can fail on a new flower. A fit that had been allowed to see the label would be answering a different question. The check is whether the geometry of the measurements recovers a grouping that the algorithm was not given. It is not an error rate for a future flower.