Same way they do in Bitcoin. Spending a double-spend before it is recognized as a double-spend.
Only derivative transactions become invalid, not all chained transactions, that's an important point.
Two partitions may not be known to each other and then later they are, because there is no global LCR forcing us to know which chain is known. And we know due to asynchrony that transaction signers can't see all activity of all other signers simultaneously. The only way I see to resolve this is with centralized servers emulating a block chain.
Double-spends in DAGs are probabilistic which I don't think was reflected in your diagrams. The decisions about double-spends are not binary (yes/no). I am guessing that maybe you have not yet conceptualized the implications of asynchrony.
You can just wait for an amount of POW to be appended to the chain containing your transaction equivalent to N bitcoin confirmations. I don't see the difference.