Robustness

This entrypoint pairs the two layers of TorchLean's robustness vocabulary:

• Robustness.Spec gives scalar-polymorphic definitions over shape-indexed tensors: tensor norms, distances, adversarial robustness, certified robustness, Lipschitz continuity, and contractions.
• Robustness.Runtime specializes those definitions to Float and provides finite, empirical diagnostics for examples and command-line checks.

We keep these layers separate on purpose. The spec layer is the mathematical language used by proof developments; the runtime layer computes observed quantities from finite samples and does not claim certification by itself.

References:

• Szegedy et al., "Intriguing properties of neural networks", ICLR 2014.
• Goodfellow, Shlens, and Szegedy, "Explaining and Harnessing Adversarial Examples", ICLR 2015.
• Wong and Kolter, "Provable defenses against adversarial examples via the convex outer adversarial polytope", ICML 2018.