Embedding model migration zero-downtime

An e-commerce semantic-search system has 300M product embeddings serving live traffic, and the ML team wants to upgrade the embedding model to a better one. The new model produces vectors in a different space, so old and new embeddings are not comparable — you cannot mix them in one index. You must migrate 300M items to the new embedding without taking search down and without a window where results are garbage. Design the embedding-versioning and zero-downtime re-embedding migration.