On construction of the collection, we compute an unsigned long which is effectively 64 boolean flags, each representing the presence of a key string of a particular length (mod 64). When reading from the collection, we can exit early if the key being tested does not map to a bit which has been switched on by the original compuation. I believe this has similarities to how Bloom Filters work.
This adds a relatively small cost on creation of the collection as small cost to each read operation. However it can speed up reads with certain data patterns especially when the difference between the maximum and minimum key length is large but there aren't many different lengths.

It is probably an opinionated change which isn't right.