add-shape-types-to-torch-model

You are porting a PyTorch model to use pyrefly's tensor shape type system.

This usually works. The example corpus contains dozens of substantial, explicitly scoped ports of real open-source research models, including transformers, convolutional encoder-decoders, autoregressive architectures, and dynamic module stacks. A first checker pass with errors or gradual tensors is a diagnosis, not evidence that the model is unportable. Work module by module and expect a few short probe–fix–check iterations before accepting a boundary.

When subagents are available, use them for bounded, independent read-only work: map the source's stated shapes, audit the model's ops against the stubs, or cluster checker errors by root cause. The main agent still owns the port and integrates the evidence; do not have several agents edit the same file.

The gates below prevent the common failure mode of writing the whole port from reasoning and checking only at the end. Apply their rigor according to the deliverable: