Creating a time-series collection on a sharded cluster forwards the command to the primary shard using the user-facing view namespace (which is inherently unsharded), meaning that even if a sharded buckets collection already exists that shard versioning information is not included in the request. When creating the buckets collection on the shard, we skip shard versions checks when acquiring the collection and then return early if the collection already exists on that shard. There is a shard version check after that, but at that point we already know that the collection doesn't exist on the shard (plus the shard version information provided with the request is for a different namespace, as previously noted).