good reasons why "critical infrastructure" is not a very useful phrase regarding the decision "who is important enough to gain a special case".
I agree that this particular term could be optimized because it actually calls for a beauty contest. But chosing the title can be postponed.
ccTLD/gTLD DNS service works very well using provider aggregateable address space. Changing glue in the root zone is heard to be needlessly difficult, but that's not something thats ideally solved by changing the address allocation policies (to avoid having to change glue records at all).
Andreas' mail included a pointer to the Vixie/Kato draft which explains why you can't increase the number of NS or A or AAAA RRs for TLD (and other) beyond a certain limit. The reasoning is similar to the root server case. So, anycasting is Best Current Practice. But a policy change that indirectly "allowed" a separate route per anycast address would sooner or later have to decide who's going to do anycast or who's doing critical vs non-critical anycast or would miss the goal because routes would again be filtered. Even if special anycast addresses would be defined, what's the threshold, i.e. how many instances would be required and who's going to check and monitor these requirements? How big is the problem currently? -Peter