mnist-classifier (champion v2)

A deliberately trivial MNIST conv_net -- the champion selected by an Optuna search and registered in MLflow, exported here as TorchScript and served by revision from this repo.

• Test accuracy: 0.9866
• Expected calibration error (ECE): 0.0023
• Hyperparameters: lr=0.0013292918943162175, conv_channels=16, dropout=0.07800932022121826, hidden_dim=128
• Source MLflow run: 4863fc86bb854bd4a8644dd3d49fc216

Preprocessing: grayscale, normalize with mean 0.1307, std 0.3081 on a centred 28x28 digit. See the write-up for the full MLflow/Optuna story.