My hypothesis is that the difficulty of writing automated end-to-end tests where failure can reliably indicate actual brokenness, combined with some of the othe

My hypothesis is that the difficulty of writing automated end-to-end tests where failure can reliably indicate actual brokenness, combined with some of the other factors I mentioned is the reason for all the focus on unit and shallow integrations that you mention.