openai/gpt-chat-latest → B (3:2)

Side B adds substantial offline tooling improvements: it introduces `compile --ingest` to replay prior events and compile a single ingest from a log, makes `scan` a fast parse-only pass instead of replaying the full reducer, and surfaces detailed `parse_error` information in both data structures and CLI output. Side A implements an SSE-based fetch flow with UI refactoring and async coordination, but much of the patch is infrastructure and code movement around an existing fetch feature, whereas Side B delivers durable developer-facing functionality and performance improvements.