Two-model loop for reviewing interactive CLI UX

To improve UX of interactive CLI programs: have a small/cheap model role-play a naive first-time user and run the binaries blind (no source), then have a stronger model triage its findings against the source. The small model surfaces real gaps (undocumented input ranges, cryptic errors) but also false positives — e.g. it typed space-separated numbers on one line and misread buffered scanf consumption as state leaking between menu operations. The strong model verifies each claim against the code and rejects the false ones before you act.