Visual Testing System
Rebuilt the visual regression suite from scratch. Replaced a 12-minute CI job with a sub-60-second Playwright pipeline and tripled coverage in the process.
12 min → <1 min, 3× coverage
What went wrong
First version tried to diff every pixel. Screenshot noise from antialiasing killed the signal-to-noise ratio. Switched to element-level snapshots and the false-positive rate dropped to near zero.

