proposition
The same index, ensured on demand instead of at startup.
The catalog above can only register what it can see WHEN IT IS BUILT, and a host that starts without an embedding model has nothing to register. Its own note said the catalog would be rebuilt on the next boot — true while such a host restarted after a model was configured, and no longer true of one that applies a provider key without restarting. The index was then never made: the model came alive, was used, and every read failed with "There is no such vector schema index".
A catalog cannot answer it either. SchemaCatalog holds materialised specs and SchemaManager snapshots the catalog beans when it is built, so a later enforce() applies the same empty catalog. This goes through PersistenceManager.indexes, which is the public way to converge one index without a catalog at all.
The host decides when — see PropositionVectorIndexConvergence.