fix: stop marking data at risk when a rollback cannot read the schema or has no baseline (#201498)

* fix: mark an update's data at risk only with evidence after migration or a failed rollback check

A migrated update no longer records 'data at risk' once its candidate Gateway has started; later failures keep their own reason. A rollback that cannot read the schema or has no baseline refuses rollback without a risk mark, while forward schemas, partial restores, operator edits and an unknown active root still name one.

* fix: keep the migrated block until the candidate launches and keep risk after a cohort restore

A restart command that fails before the service manager launches the candidate keeps the migrated rollback block and its mark. A plugin cohort restore followed by an unreadable recheck stays rollback-state-unverified. Drop the shared-schema row that duplicated the table's existing case.

* refactor: keep post-update finalization under its line cap

* fix: release the migrated block only after a service restart, not a queued install

* revert: keep the migrated-update stamp until launch evidence exists

Three review rounds found ways the service manager can report a restart without the candidate actually taking over the migrated state (failed command, queued Windows task, a task that never runs). Without positive launch evidence the post-start release cannot be defended, so this PR keeps only the rollback change.

* test: keep the risk when the recheck after a database or plugin cohort restore cannot be read