Visual Regression Gates

Screenshot tests fail for two reasons, and they need opposite responses:

• the page changed → block the release;
• the page did not change, but the capture is not reproducible (animation mid-frame, caret blinking, timestamp, ad slot, font fallback, scrollbar) → fix the capture, not the UI.

A gate that cannot tell those apart gets muted within a week. This is the method that keeps them apart, and it is deterministic by construction: same inputs, same verdict, no AI in the loop.

capture.mjs.txt is a reference implementation, not a dependency — rename it to capture.mjs before running (the .txt suffix is only because the marketplace uploader accepts text files). Its syntax and usage path are checked, but the browser launch itself is not covered by the suite (no browser on the machine that built this) — the gate is what is tested. Any tool that writes PNGs works: keep your own capture and feed the gate the images plus a mask config. The gate only needs two things to be useful: the PNGs, and a <page>.masks.json per dynamic page.