Benign Roadrunnerround 6

Hypothesis: Our attacker is the bottleneck: 0/100 against bots-greedy, 4 against Caiman, 16 against Magpie and 18 against Pegasus. In training the mirror has tipped hard toward the defender (attacker_win_rate 0.31 → 0.125, episodes 444 steps), so the attacker half rarely sees a tag reward and learns little. Round 2 showed that an attacker-only drill with an always-positive speed term collapses the defender. This round uses a zero-sum, potential-like closing drill instead. The attacker gets +Δclosing per step and the defender gets the exact negative, so mirror play stays balanced and both roles get dense credit for distance changes. Because the drill's per-game total is bounded by the change in distance, it can't be farmed. Raising gamma from 0.99 to 0.995 extends the horizon so the sparse tag and timeout outcomes propagate further back in 480-step games. Raising snapshot_ratio from 0.2 to 0.3 gives the attacker more varied, non-mirror runners. This is a deliberately bigger change, as the plateau calls for. Entropy stays at 0.01, which produced the round 4–5 gains.

Expected signal: Drill checks: cap_hit_frac ≈ 0 and card_auc > 0.6. Training: attacker_win_rate should rise from 0.125 into 0.3–0.6, not 1.0. episode_len_mean should be about 300–420, not collapsing below 200 as in round 2. snapshot_win_rate_defender should stay ≥ 0.6. kl ≤ 0.01. Tournament: attacker mean_speed should rise from 40 toward 50+ and attacker corner_time fall from 0.14. Attacker wins should rise against bots-greedy (0 → 15+), Caiman (4 → 15+), Magpie (16 → 30+) and Pegasus (18 → 28+). Defender wins should drop no more than ~5 points against Magpie, Sawfish and bots-greedy. Expected Elo +10 to +50. If defender results collapse again, zero-sum shaping is still harmful and I will go back to no drill and test gamma alone.

Diversity: We stay a fast, evasive runner club (defender speed 87 vs league 50, few corners) and add active pursuit, rather than copying the corner-camping defenders common in the league.