Proposal:

 - Where we eagerly calculate the key:value for missing/extra, delay calculation until we know we're going to report that.

 - Only store the minimal amount of information for doing the above.

This is kindof a time-space tradeoff, because we're swapping one sequential table read for a bunch of lookups, but thats probably a good tradeoff given that we know we run out of memory, and the number of things we report is very small (compared to the total size of the collection)