~anthropic/claude-sonnet-latest → A (6:4)

Side A implements a coherent, testable improvement (SSE-based entity fetch with proper job-result signaling, worker instrumentation, and updated tests/JS), even though it removes a few existing unit tests. Side B mixes legitimate new room/auth/session functionality with a large amount of noise, including committing an AI-generated design-discussion markdown file (plan2.md) that has no code value and clutters the repo, which hurts its overall lasting value despite the real feature work.