We need to classify errors by return code

Currently, the client doesn't usually raise exceptions. Thats ok, but complicates later parsing in the txn code, to determine whether an error is transient, etc... Perhaps client exceptions will make that easier - so the understanding of rc need not leak out of the client.

Or, we expose the return codes (symbolically), which is ok too. The thing is, we can't have comparisons with actual integers – that is fragile and unreadable/unmaintainable.

Up until now, we had bigger problems. But now that we need the retry logic to work correctly in the tests, this needs to be addressed. Four of the 5 fails left in StandardTest are due to retry logic in insert, which is tricky enough without comparisons involving magic numbers.