openai/gpt-5.3-chat → B (9:1)

Side B makes a deep architectural change: it removes `EntityStore` and the `EntityImported` event entirely, rewrites projection application to exclude payload persistence, and introduces ephemeral caching with TTL eviction (`put_ephemeral_content`, `evict_content_older_than`) plus updated tests. Side A mainly adds a planning document and a thin `RouteContext` wrapper delegating to `ThreadNav`, with minimal functional impact.