QwenJev v0.4.3: 25-public-games eval in-repo (run_eval incremental report, fetch_games, CPU feasibility verified)

v3: real-games training round tooling

real-game Stage B trained checkpoint: 120 steps, 25 games, 4.90GiB steady

QwenJev v0.4.3: 25-public-games eval in-repo (run_eval incremental report, fetch_games, CPU feasibility verified)

real-game selfplay data: 25 games x 3 eps, 8863 steps + 20414-sample dataset

QwenJev v0.4.3: 25-public-games eval in-repo (run_eval incremental report, fetch_games, CPU feasibility verified)

collect_selfplay: ARC_ENVIRONMENTS_DIR env-var for offline game files location

QwenJev v0.2.0: vision tower on the System-1 inference path (tiny/hf backends, mock env, e2e local run)

QwenJev: multimodal-retrofitted NanoJev for ARC-AGI-3 (initial skeleton)