The upgrade reverts when the pixels disagree

The agent upgrades the CSS framework overnight: two majors, a codemod, forty-one files. The tests pass. Tests don't look at anything.

So before touching a file, the agent snapshots the container and captures a baseline: GET /api/v1/browser/screenshot for each of the twelve key pages, full-page. After the upgrade it captures the same twelve and compares. Layout holds on eleven; the pricing table collapses at mobile widths.

It doesn't file a ticket about the regression and leave the damage running. It restores the snapshot: upgrade gone, machine back to the state that matched its own baseline: and leaves a report with both screenshots side by side.

Green tests mean the logic held. Matching pixels mean the product did. The upgrade reverts itself when the pixels disagree.