Misreading a t-SNE Plot

An analyst runs t-SNE (perplexity 30) on 1,536-dimensional text embeddings of 20,000 support tickets and shows you a 2-D plot with about eight blobs. They conclude:

(a) "The two blobs at the top left are closely related topics because they are next to each other, and the blob at the far right is a completely unrelated topic." (b) "The big blob is the most common ticket type; the tiny blob is a rare edge case." (c) "I will save the 2-D t-SNE coordinates as features for a ticket classifier."

Which conclusions are unjustified and why? Then describe a more defensible pipeline for exploring and using these embeddings.