Our CI is red because of a litellm drift: our tests are assessing values against a live metadata litellm db This PR ships an immediate fix This PR could fix the supply-chain issue. But it's not clear we should fix it: models metadata and capabilities change every day. Do we really want to manually sync a copy of that file in our code? (could be an automation though) For metadata and capabilities (not super sensitive) we could trust litellm?

approved the first one for now