openai/gpt-chat-latest → A (3:2)

Side A introduces substantive architectural improvements to the running server: it adds a settlement worker that batches vote persistence and ranking recomputation, introduces cached ranking reads (`ranked_items_cached`) so HTTP handlers can use read locks instead of write locks, and removes an obsolete demo counter feature. Side B adds a large new `durable` workspace crate and streams event-log replay while moving raw entity payloads into RocksDB, but much of the patch is dependency, documentation, examples, and vendored infrastructure, making the immediately integrated project impact less concentrated than A's core runtime and concurrency improvements.