Three decades of code similarity detection can be told as two changes in the unit of comparison, from lines to token sequences to syntax trees. Each shift shrank the review queue by an order of magnitude. Here’s what each era bought, what it still misses, and how generated code forced the field to add a second axis.