Each container job now looks up the image's index, manifest,
and config digests from the registry,
keys an `actions/cache` entry on the config digest,
and only uses a cached image if its ID matches one of those digests;
otherwise it is discarded and the image is pulled (with retries) instead.
A poisoned or stale cache entry can therefore never be used,
and a cache hit needs only a couple of small manifest requests,
so this scales regardless of pull rate limits.
