LateArrivingModels

Declares that a named provider's models may appear AFTER startup.

Implemented alongside RoleResolver or EmbeddingRoleResolver, or by a LocalModelCatalog, whichever object knows. It says nothing about how a role resolves; it answers a different question, which only startup validation asks: is a role naming an unregistered model under this provider a typo, or a model that has not arrived yet?

Without it the answer is "typo", and it is fatal - correctly, for a deployment that holds its models, because letting it start moves the failure to whichever unrelated call first asks for that role. A provider whose models are pulled on the host is the exception: roles.cheapest.docker naming a model nobody has pulled yet is the ordinary state of an appliance before setup, and refusing to start is refusing to reach the point where the operator could pull it.

This is why ConfigurableModelProviderProperties.embeddingRoles is not checked at all - every entry there may be for a provider this process cannot serve - and it extends the same tolerance to the ONE case the chat check still treats as fatal: an entry under the same provider the default LLM comes from.

Only the PROVIDER-QUALIFIED maps are covered. The flat llms and embedding-services maps name no provider, so nothing there can be attributed to a runner rather than to a typo - which is the reason a model that may arrive late belongs in a provider column.

Inheritors