Version Resolver speed training

25/25 correct; first PB 11.767s→4.307s. Faster than full filter/scan: cache only 2.2.x and 2.1.x, check exclude; if 2.2.x is not excluded, emit it, else emit 2.1.x. Ignore 1.9.9/3.0.0; fall back to full semantic filtering if generator shape changes. Bottleneck: runtime/tool variance.