Adaptive topology experiment — prototype

A second, separate isolated A/B test: does letting the network's own WIDTH (a real, learnable per-neuron gate — any feedforward neuron can learn to switch itself off) and CONNECTIVITY (a real, learnable feedback loop — the network can learn to revisit its own first pass through a scaled addition back into its own starting input) lower held-out loss, versus today's fixed-width, single-pass architecture? Trains two SEPARATE, disposable models on identical real data — never touches Language Playground, Chat Model, Chat Model Lab, or the separate APL activation experiment.