openai/gpt-chat-latest → A (20:1)

Side A adds substantive functionality and architecture: it introduces fast non-replay scanning, richer parse error reporting, a new `compile --ingest` workflow that reconstructs prior reducer state for a single event, improved CLI behavior, and tests/documentation to support these features. Side B only deletes `test/runner.clj`, removing a test entrypoint without adding new functionality or evident long-term design improvements.