He's saying that if you explicitly separate your crates out, and then you accidentally introduce a circular dependency you will get a compile time error. But if you have one big crate and rely on the compiler to detect internal dependencies for performance, then when you accidentally introduce a dependency that you didn't mean to it will happily compile but it will be slower so you probably won't notice and it will be really hard to debug.
Doing all this tracking correctly though across layers of compilers is tricky if you want to shoot for absolute optimal performance. At the limit for truly optimal behavior, you'd track all the input information the compiler used when generating code and map it back to the source level (i.e. did I inline function A into function B - then I need to regenerate function A if B changes, but if no inline then don't regenerate A), which can be hard as most compiler optimization passes are information destroying. So yeah, I don't doubt that whatever "practical" middle ground is found as a realistic implementation can be tricky to reason about, but I suspect it almost doesn't matter if done well because it'll always outperform what you would have done by hand as outlined in this post / do it for you for free in 90% of cases.