Two more quality passes on the production side.

Full production manifests are no longer embedded into memory. They were being written to the vector store as-is on every finalize, which made retrieval noisy (a manifest is too big to be a good semantic hit) and cost real time on the write side. Skipping the embedding step keeps memory tidy and shaves time off every finalize.

And the non-film production workflows (novels, audiobooks, music, YouTube, and social) picked up their own optimization pass, matching the film work we shipped this morning so every content type finalizes on the same fast rails.