The per-job caches, I believe, fully override the globally defined caches.
This presumably explains why the cypress jobs weren't cacheing after the last cicd update.
Fixing this should further reduce the number of yarn.lock builds per pipeline run.