For instance, there seemed to be a lot of "hash collisions": vocabulary words that I kept confusing with one another. Some of these were clearly phonetic: hai (shark) and kai (probably).

That problem is called memory interference. I think reading Wozniaks 20 rules, gives you a good elementary understanding of concepts like that.

In general there doesn't seem to be a good way to predict memory interference in advance.

When faced with apparent interference I usually make a card specifically for the interference:

Front: (kai / hai) -> shark
Back: hai

Front: (kai / hai) -> probably
Back: kai