+ * 4. Elimination of LCS computations that will never breach the configured
+ * threshold. Two measurements are used as rejecting filters (i.e. a failure to
+ * exceed the threshold prevents further measurements being made):
+ *
+ * 4a. Comparsion of string lengths: String lengths must be within given bounds
+ * to satisfy the user-supplied similarity constraint. A negative result avoids
+ * invoking the O(m * n) behaviour of LCS at the cost of O(m + n) in the two
+ * strlen() invocations.
+ *
+ * 4b. Comparison of character distribution: Cosine similarity[2] is used to
+ * measure unordered character similarity between the input strings. The
+ * implementation is again O(m + n), and avoids the O(m * n) behaviour of LCS.
+ *
+ * Performance will vary not only with the number of input strings but
+ * with their lengths and relative similarities. A large number of long input
+ * strings that are relatively similar will give the worst performance.
+ * However to provide some context, strgrp can cluster a real-world test set of
+ * 3500 strings distributed in length between 20 and 100 characters to 85%
+ * similarity on a 4-core 2010-era amd64 laptop in approximately 750ms.
+ *