If uninitialized replicas do not end up being initialized (e.g. because the snapshot was cancelled or failed), there is currently no mechanism to garbage collect them. These replicas currently have a non-negligible cost (see related support issue) -- they're ticked (#73362) and also prevented from quiescing (#73397) -- but even when those costs are dealt with we should make sure they're removed eventually. If nothing else, to avoid the confusion of having them show up in range reports and such.

There's a stale PR for this in #47982, and further discussion in #73362 (comment).

Jira issue: CRDB-11582