They could do it without recompilation, but something like changing the obfuscation and recompiling for every copy would likely make it much harder to get rid of the watermarks even if you can compare several different copies

(though they could also have multiple watermarked sections so that any group of for example 3 copies would have some section that is identical, but still watermarked and would uniquely identify all three leakers. The amount of data you need the watermarks to contain goes up exponentially with the amount of distinct copies, but if you have say 1000 review copies and want to be resistant to 4 copies being “merged”, you only need to distinguish between 1000^4 combinations so you can theoretically get away with a watermark that only contains about 40 bits of data )