the record had become the largest thing in it

Large generated tables had been committed alongside the code that produced them, repeatedly, over months. Nothing about any single commit was wrong. The aggregate was: the repository had grown to gigabytes, most of it result artifacts that could be regenerated, and every clone and every backup paid for all of it.

Removing the files would not have helped, because the history keeps them. So the decision was to rewrite the record itself, which is the closest thing to an irreversible operation available here. Every branch, every working tree and every unpublished line of work has to be accounted for before the first step.

The execution was sequenced rather than brave. The growth was attributed to named sources first. A dry run was measured before the real one. A freeze was posted, the branches with unpushed work were woken and settled, and the freeze was respected before anything was rewritten. The result took the repository to a small fraction of its former size, and the working trees were reset rather than left pointing at objects that no longer existed.

Then the half that lasts: a mechanical guard at the boundary. Pushes that would introduce an oversized file are refused before they land, tested before it was relied on, installed through the mechanism the tooling actually honours rather than one that can be silently ignored, and paired with a retention rule so the guard cannot be defeated by deleting what it protects.

What it proves is that the team will pay for a costly correction once and then a small amount forever to avoid repeating it.