GraphSpec Primitive Packs

This umbrella re-exports the non-core GraphSpec primitive packs.

GraphSpec primitives are operation adapters, not models:

• a primitive wraps one reusable operation, such as convolution or pooling;
• a model composes primitives into an architecture, such as cnn2 or ResNet18.model.

The small always-available primitives (linear, relu, softmax) live in NN.GraphSpec.Core because they are part of the minimal sequential DSL examples and lowering interface. Larger domain-specific packs live under NN.GraphSpec.Primitives/* so the core language does not become a grab bag of every layer TorchLean supports.

Current extension pack:

• NN.GraphSpec.Primitives.Vision: CHW convolution, max-pool, flatten, BatchNorm, and global average pooling adapters.

Adding a new operation to GraphSpec means giving it both meanings:

• pure Spec semantics for proof/reference use;
• TorchLean program semantics for execution/training.