openai/gpt-chat-latest → B (4:1)

Side B fixes substantive feed correctness by introducing durable ingest-order anchoring instead of timestamp-only cutoffs, preserving permission checks, adding room metadata to feed responses, and covering multi-user/private-room edge cases with comprehensive integration tests. Side A mainly refactors ranking from contributor-level to commit-level with UI/evidence updates and tests, which is a meaningful feature but less fundamental than correcting feed consistency and visibility bugs.