Self-supervised learning theory

This is the umbrella import for TorchLean's finite self-supervised objective semantics.

The chapter contains:

• masking primitives for patch/token objectives;
• predictive-view SSL objective algebra;
• MAE-style masked reconstruction losses;
• JEPA-style joint-embedding prediction losses; and
• VICReg/Barlow-Twins-style finite collapse guards.

This layer is intentionally modest: these files formalize the objective-level contracts that model code should preserve, rather than claiming full optimization or population generalization guarantees. That keeps the statements reusable across ViT, convolutional, Mamba, or custom backbones.