Optic-lobe steering

The shipped browser model is a documented leaky rate network over the MaleCNS optic lobe: tau_i dx_i/dt = -x_i + wScale * sum_j W_ij r_j + ext_i + bias_i, with per-cell-type tau and bias taken from flyvis and a rectified rate ceiling. Not L2: the step function reads the weight array and never writes it, so nothing changes synaptically while the fly is steering. The repo does contain PyTorch training (Adam, loss.backward()) that fits log_tau, bias and a per-cell-type-pair log_strength, but this is the most common near-miss on the scale rather than plasticity: it is offline, it optimises a few hundred type-level parameters instead of individual synapses, it is driven by a gradient rather than a modeled neuromodulator, and it bakes fixed values into the weights the browser then loads and holds constant.