sorry I thought I replied yesterday but apparently my comment got lost.

Using a (hash)set instead of a bloom filter caused memory usage to explode, which in turn led to memory thrashing/swapping which in turn led to slow processing times.

The rest of the code runs in constant time (well, linear in the size of the item, with num(items) >> size(items[n]) ) and it did work fine for smaller inputs.

How much slower? Just how much wall clock time did you lose?

hard to estimate I'm sorry. The first naive implementation had been running for more than 15 minutes before I thought it as too much and killed it. Rewrote version run all in about 5.