roombaTraining.module

Measured motor-rate calibration: 36 training stimulus pairs and 12 held-out pairs, sampled from the full GPU fruitFly model. Polynomial wheel decoder trained by gradient descent. Held-out MSE 0.0000286741 versus 0.00827995 for the identity decoder. This trains the motor readout, not cleaning decisions or connectome synapses. Raw measurements are embedded in init.luau and archived in calibration.json.