By this reasoning, though, there is no such thing as localized harm. The reason I can't abide by the idea that "there's no such thing as localized harm" is that, when you actually try to analyze nonlocal harms caused by personal decisions, you get swallowed up by the butterfly of doom.

It's the butterfly effect. For example, a lot of software that actually gets written is a net negative to society, even if it functions perfectly. So does making it more efficient actually benefit anybody? And a lot of other software is embedded in organizations that will add features to the software until it fails, expanding like an ideal gas to fill whatever space it's given, so even if you make it more efficient and less failure-prone, you're only really delaying the inevitable anyway. However, making a bureaucratic organization less efficient might not actually stop it; consider, for example, how the Social Security Card was originally engineered to be unusable as a national ID, but got used as one anyway, so now the United States not only has a national ID that most citizens didn't want, but we're stuck with a bad one. However, identity theft might actually be considered just another case of externalities, and if the bureaucrats had to eat the cost of easy-to-forge national IDs, this problem might have gotten fixed.

I think you can analyze nonlocal harms, but not using informal reasoning in a chatroom. There are too many possible interactions in the real world to fit them all in your head. You end up with an impossible-to-analyze infinite regress.

Instead, nonlocal harms should probably expect real-world measurements to prove that they actually exist and aren't entirely being washed out by the much larger effect sizes of unrelated phenomena.

Externalities is a concept in economic theory. It shows how net negative behaviour occurs, even when all actors act perfectly rational and have perfect information (while optimizing for their own gain). Bad software does simply not map to this concept in the same way environmental damage does. In your example people use bad software, against their interest, despite better alternatives.