That’s what I always wonder too. If we assume an attacker can change the contents at example.com/app.zip, why should we assume the hash published at example.com/download.html is any more secure?

In some setups, you're required to trust both the maintainer and the mirror, which are not always the same party. If someone can generate a collision, it means a mirror can mount an attack even when the maintainer is corrected trusted.

You can pass review that way. You publish a clean artifact that gets reviewed and vetted, and for the actual attack you replace the vetted artifact by the bad one. If you trust MD5, or as the article shows, even a good 128 bit hash, like truncated SHA-256, you get pwned. That's why you don't accept MD5 based signature as well.

People that take security seriously enough to check hashes should not trust MD5 so the scenario is not super credible, but people still publish MD5 hashed like it's the early 2000s.