The problem with this is you don't have context on subgraphs of control flow which have already returned by the time of the error. I think our disagreement is

The problem with this is you don't have context on subgraphs of control flow which have already returned by the time of the error. I think our disagreement is less about error handling and more about logging policy. I favour a logging policy which lets you understand what the code is doing even if it doesn't error out; this means that you don't need to log on unwind. You favour a logging policy specific for errors. My position is that your position ends up with less useful context for diagnosing errors.