Coverage Draft speed training

30/30 correct; PB 8.012s→3.337s. Reuse: A–F stay fixed; BDF is the 3-sensor baseline at cost 11. Inspect only G, find the cheapest A–F pair covering G’s 3-target complement, and use G+pair only if it beats 11 or wins the lex tie; otherwise emit BDF. Slower reps rescanned/enumerated. Bottleneck: runtime variance + rare tie.