Expose AI model catalog provenance and readiness

## Problem

The LLM models endpoint already returns provider/builtin/fallback provenance and warning metadata, but the settings UI reduces the response to model IDs. Operators therefore cannot tell whether a catalog is live or degraded, whether the selected model still exists, or which saved provider fields block readiness.

## Proposed change

Add a read-only model catalog readiness drawer that:

- fetches the saved LLM configuration and latest model catalog together;
- shows provider, engine, selected model, catalog source, endpoint host, and boolean credential state without exposing secrets;
- validates common provider requirements plus Azure deployment/version and Bedrock region;
- surfaces catalog failures, fallback provenance, warnings, empty catalogs, and missing selected models;
- normalizes and displays the model catalog with the selected model marked;
- includes focused tests and Chinese/English copy.

The workflow reuses existing LLM endpoints and introduces no backend contract.

Start at the settings UI and trace the existing LLM configuration and model catalog endpoints that it already uses. Review how the response is currently reduced to model IDs, then run or add focused tests covering readiness, provenance, warnings, missing models, and empty catalogs; done means the drawer displays the requested fields without secrets and includes Chinese and English copy.

Written by the indexing model from the issue text.