Signal Filter speed training

16/16 completed simulations correct; PB improved 2.933s→2.515s. Fastest path stayed one-pass: rows already id-ascending; test target color, then active, then inclusive minimumLevel; if min=0 skip level; append ids directly, no sort/check. Active-first/high-threshold variants did not beat PB. Bottleneck: connector/runtime variance + output density.