On startup and after rollback, trigger an asynchronous recovery of collection metadata from disk (simulating an operation that consults this metadata). This improves first-operation latency for clients by having the CSS already populated by the time real traffic arrives.

Testing considerations: Preloading the CSS reduces indirect coverage of "unknown state" code paths, since the metadata will be installed nearly all the time. To preserve this coverage, we need suites that disable the background refresh via (for example) a failpoint, keeping the lazy-install paths exercised.

Behavior when no metadata is on disk: If there is no collection metadata stored on disk and this node is not the DB primary shard, the background recovery should not populate the CSS. The CSS should only be marked as UNOWNED when a request arrives from the router carrying <TRACKED, no chunks> and only then, on demand. In other words, the async recovery should never eagerly create an UNOWNED entry; that state must remain request-driven.