At any time either party can close out the channel unilaterally by broadcasting the most recent state. The person that broadcasted it gets refunded after a delay (assuming the uncooperative case where the other party in the channel goes off the Antarctica or something -- in a cooperative case, both parties close out immediately with the current balance).
Leaving coins in LN is fully backed on-chain, of course, since these are real bitcoin transactions passed around. The added benefit is you can transact instantly in high volume, so coins in LN will probably have more use than coins off LN when it comes to payments.
Say Bob says I want my $10 to stick on OKCoin now. He broadcasts that he wants to close the channels for all those thousands of people now? Doesn't that mean the transactions go through on the actual network costing him ~$0.05 for each $0.01? Or am I understanding settlement completely wrong?
You update your channel to tell Bob, I will pay you conditional upon a proof that you paid Carol. Bob pays Carol and provides proof. If Bob didn't provide proof after an agreed timeframe, this gets cancelled. There can be an arbitrary number of people between you and Bob (7-degrees of Kevin Bacon, but probably don't go too high!).
This is done completely off-blockchain. At any time this can be moved on-blockchain, but since the blockchain will enforce the payment as part of its transaction scripting language, you both elect to do it off-blockchain. Think of this with the system intent/design where it represents an enforcible caching layer where invalid cache can be corrected on-blockchain.
You can do thousands, even millions of these transactions off-blockchain. Only the final balance is net-settled on the blockchain. So you finalize your balance between Alice and Bob if you want to close out the channel. Note that the channel between Bob and Carol could still be left open and everything still works fine!
Since it's scalable in a routed graph (think switches instead of hubs), you could potentially do millions to billions of transaactions which route value, much like routing packets on the internet.
If you have any questions, do send me an email. I'm working on making the explanation better so feedback is always appreciated.
Say I want to pay Bob $0.001. That's too small for a bitcoin transaction since the fee is higher than that. So from what I understand I would "load" say $10 into the lightning network then create a transaction over the network to Bob for $0.001. Say tons of other people do this and now Bob is sitting on $10. How does he turn that into actual Bitcoin? Like get it out of the lightning network. Doesn't he have to close off all those thousands of channels?