~x-ai/grok-latest → B (2:3)

B fixes a real reducer correctness bug: zero-ratio votes were rejected only after ensure_item and voted_pairs.insert, leaving ghost items and false voted pairs; moving the guard before side effects (with an updated test) protects ranking state integrity. A is solid tooling work—fast parse-only scan, richer parse_error surfaces, and compile --ingest—but it improves offline ergonomics rather than core simulation correctness, so it trails slightly on lasting project value.