If network connectivity is severed, then the system can't be used reliably, obviously. This is much easier to detect with Ripple than it is with Bitcoin -- you never consider a transaction validated until you see sufficient validations for a ledger that contains it (or is chained from a prior ledger that does), which you can't if the network is severed. With Bitcoin, you'd have to notice that the block discovery rate seemed strangely low, which you can only infer statistically*. With Ripple, you'll notice the absence of validations from all nodes you're disconnected from immediately after the next ledger closes.

You are correct that we can only simulate robustness in the face of attack models that we can think of. Any system is potentially vulnerable to attack models that weren't imagined by its designers. This is another reason we want to get the system completely open and decentralized as quickly as possible. We don't want people to have any reason not to trust the system.

* Until and unless someone implements more clever solutions, I believe some people who accept zero or one confirmation Bitcoin transactions do implement such solutions now.