Planning an Embedding Model Migration

Your company's embedding provider is deprecating the model you've used for 18 months in favor of a newer version with a different vector space and different dimensionality. Your index holds about 40 million chunks across several products, serving live retrieval traffic 24/7.

1. Why can't you just start writing new chunks with the new model and leave old chunks as they are?
2. Design the migration so retrieval quality never regresses for users during the cutover.
3. What would make you decide to delay or roll back the migration?