The problem is not so much "when unsafe code gets called indirectly by safe code", which is fine if the unsafe happens to be sound. The problem is that "when C-like or Zig-like unsafe code wants to call safe code", the safe code is running in a context where its implied invariants may be violated, leading to insta-UB. Hence code that's intended to provide "library-like" facilities to unsafe blocks cannot be idiomatic Safe Rust, and it needs to be written even more carefully.

For instance, any &mut reference in Rust is assumed not to be aliased, and any &reference not involving UnsafeCell<...> is assumed not to be written to. These implied contracts can be loosened, e.g. by using &Cell<> if applicable (may alias, can be read or written safely, but only "as a whole object": access to the internals does not escape beyond any single operation) which is arguably closer to idiomatic C.

MaybeUninit<> is another common example: C and Zig code often works with possibly-uninitialized data, but this possibility has to be accounted for explicitly in a safe Rust interface. It's always insta-UB if a safe Rust function is passed uninitialized data, even when the equivalent would work idiomatically in C.

This is what makes Rust Rust and I'd say what makes Rust popular. The way safe Rust is safe. Of course this comes at the expense of making writing unsafe a lot harder.

Though I can imagine that unsafe Rust still has to many of the safe Rust's rules. So there could be a better unsafe language that has fewer restrictions.