Encountered in relation to #2566, specifically the new 200 requests/minute GHA rate limit for cache writes.
With concurrent builds running on different VMs, we've found that it's not uncommon for the initial cache read/write check to occur during a window where the service is rate-limited. When this occurs, the cache is failed over into read-only mode, which then persists.
RateLimited errors are by definition recoverable, and so this case should ideally be handled differently.