The multilingual-status example is a good entry point: using a display string as a shared invariant is the kind of mistake that feels harmless until it isn't. What I found most interesting is the holonomy framing — that the same connections can demand different repairs depending on how corrections propagate is a genuinely non-obvious insight. And the distinction between knowing a repair exists (0 queries to decide) and actually constructing it (2 queries) is the kind of result that should shape what we let agents change unsupervised.