"but given the law enforcers say what you are doing is not allowed, I can only assume that it is, indeed, not allowed"
The law enforcers so often wrong about what the law does and does not allow that somebody had the foresight leave that judgement up to another branch of government entirely.
Stonebraker says real systems can obviously achieve "CA" when one node is partitioned off. But he slyly switched out the theorem's definition of "available" for a more colloquial definition. The CAP theorem definition requires every live node, including the lone segregated guy, to respond to clients. Real "CA" systems don't do that--they're "available" because the non-partitioned nodes are still responding.
There is a tradeoff though--accepting a smaller probability of success increases the variance, which deterministically drags down the returns. To see the effect, extend your argument to the extreme case of very very early valuations. You have $100k to invest, and you invest $1 each in 100k companies, each with a 1 in a billion chance of being a billion dollar company. Though your expected value is the same, at the end of the year, most likely you have no money left to invest for next year. The same effect holds in less extreme cases--if you increase your variance and lose 50% of your money, you can't just make it back by having a 50% upswing the next year--you need a 100% return to recoup it. So variance matters.
The law enforcers so often wrong about what the law does and does not allow that somebody had the foresight leave that judgement up to another branch of government entirely.