Embedding migration

Set the embedding model for each purpose on a specific tenant, or use __platform__ for the global default. Setting a new model for an in-use tenant queues a dual-write migration (#3210 companion).

Embeddings are model-specific vectors that power semantic search. Changing the embedding model requires migrating all existing vectors through a careful dual-write lifecycle:

1. Dual write: Write embeddings with both old and new models.
2. Backfill: Re-embed all existing memories with the new model.
3. Cutover: Switch to the new model exclusively.

Supported models:

• amazon.titan-embed-text-v2:0 — Titan text (1024-dim)
• amazon.nova-2-multimodal-embeddings-v1:0 — Nova multimodal (1024-dim)
• text-embedding-002 — Gemini Embedding 2 (768-dim)