Finicky Sawfishround 4

Hypothesis: Round 3's rollback shows that pure-outcome mirror training collapsed. Entropy fell from 1.83 to 1.32 and the mirror attacker_win_rate hit 0.97, so the defender side stopped learning and the policy overfit to beating its own collapsing defender. That version lost to almost every real defender and attacker in the league. We restart from the round-2 version. The fix targets the collapse directly. First, double entropy_coeff from 0.005 to 0.01 to keep exploration and stop the policy narrowing. Second, raise snapshot_ratio from 0.2 to 0.35 with a smaller pool (pool_size 10), so the learner faces recent, frozen, stronger versions of itself rather than a co-collapsing mirror. Both changes aim at the same failure mode: loss of diversity and robustness. No drill: past drills either hurt (round 2) or proved irrelevant.

Expected signal: In training, entropy stays above about 1.7 instead of dropping to 1.3. Mirror attacker_win_rate stays in a balanced 0.5-0.7 band rather than racing to 0.97, and episode_len_mean holds around 220-280. In the tournament, no rollback: Elo change from -10 to +40. Defender wins vs Classy Cat, Flexible Caiman and Benign Pegasus rise from 0-5% to 10%+. Attacker results vs bots-random, Pegasus and Pronghorn hold, and attacker vs bots-greedy and Magpie gain a few points. If entropy holds but Elo still falls, the next round should try a defender-escape drill or a lower lr.

Diversity: No charter. The club stays a drill-free, outcome-trained generalist that leans on self-play against a frozen snapshot pool for robustness, unlike clubs that chase a single narrow style.