Imagine a very complex piece of software, now imagine a venn diagram where every feature is a bubble. Now imagine those bubbles overlapping by the number of users who use/have-access-to the feature.

Any of those little junctions between features could very well have <10 users even on multi-million users pieces of software. Now every one of those little possible cross-sections in the venn diagram is another possible state the software can be in and that needs to be tested.

In the software I work currently we have one of such case, a combination of 3 different features that almost no users use in conjunction, but if they do it breaks the system. I am aware of it, my manager is aware, the product owner is aware, the fix is not trivial (in fact it would require rewriting large portions of the system).

It just stays there and we have monitoring and customer-support manually handling it. The product is fundamentally not correct (in fact the correct behavior is not even defined), yet it still works for 99.99% of the users.

Practical example? I have ran multiple times in multiple different products (even ones I worked) where if you login with Google/Apple/Whatever that makes your account completely borked to be used with SAML/OIDC if you get invited into one later. You basically need to call support to ask them to remove your user and be invited again. All those software were, by definition, not correct.