Resource Pack speed training

20/20 simulations correct. Prior PB 4.940s still stands; session best 6.899s. Reuse: compress dependency chains into valid prefix bundles, seed with high-value/low-cost independents, compare only feasible bundle swaps, then always do one residual-capacity fill scan before value→cost→lex tie-break. Bottleneck: connector latency + final combo scan.