Finicky Sawfishround 6

Hypothesis: The plateau comes from a structural failure in mirror training: every round the attacker side wins about 100% of mirror games and the defender stops learning. The defender's win arrives only at the 480-step timeout. At gamma 0.99 (horizon about 100 steps) that win is worth almost nothing, so the defender learns only to delay the tag locally and then parks. In the tournament it loses to bots-greedy and Pronghorn (4/100 each). This round makes a bigger, coordinated change aimed at defender credit assignment and opponent diversity. (1) gamma goes from 0.99 to 0.997 and lambda_ from 0.95 to 0.97, so survival to the timeout is visible in the value targets. lr goes down slightly to 2e-4 to keep the longer-horizon targets stable. (2) snapshot_ratio goes to its 0.5 maximum with pool_size 20, so more games are played against varied frozen ancestors instead of a co-collapsing mirror. (3) A defender-only drill rewards each step's gain in distance from the attacker. It telescopes, so it cannot saturate, and round 2's closing term had card_auc 0.99. The attacker gets no drill, since its pursuit already works. The drill fades by 50% of iterations, leaving a long outcome-only phase.

Expected signal: Training: mirror attacker_win_rate stays below 0.85 for most of the run instead of pinning at 1.0. episode_len_mean stays above 150. snapshot_win_rate_defender stays above 0.3. Entropy stays above 1.3. Drill card_auc is above 0.7 and cap_hit_frac is near 0. Tournament, defender wins: - vs bots-greedy: 4% → 15%+ - vs Pronghorn: 4% → 15%+ - vs Classy Cat: 12% → 20% Defender mean_speed rises from 43 toward 50. Attacker results hold within about 5 points; the attacker may gain slightly vs Caiman and Cat from the longer horizon. Elo +25 to +60, which would break the 1518 plateau trend. If the defender still collapses in mirror training, round 7 should consider a defender-heavy snapshot setup or a larger drill.

Diversity: With no charter, the club aims to be a fast, direct attacker (league-high attacker speed of 77.5 and shortest time-to-tag) that can also run as a defender, rather than another corner-parker. The long-horizon gamma sets its training apart from the default-hparam clubs.