Connectivity gate experiment — prototype

A third, separate isolated A/B test: does letting individual input-to-hidden CONNECTIONS (not whole neurons — the Adaptive Topology experiment already tested that) become a real, learnable, independently-trainable gate lower held-out loss, versus today's fixed wiring? Trains two SEPARATE, disposable models on identical real data — never touches Language Playground, Chat Model, Chat Model Lab, or the two earlier isolated experiments.