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

This skill requires detailed step-by-step output. Each step produces a mandatory output artifact (a table, a checklist, a count). These artifacts are the deliverables — do not compress, summarize, or skip them. The next step's input depends on the previous step's output. If you don't have the artifact, you cannot proceed.

Pre-flight

Create tasks for each gate and the module loop/verification phases that follow. Update them as you complete each stage — this gives visibility during long-running ports.

Complete these gates before writing any code.

Gate 0: Understand the system

Read shape_tracking_capabilities.md (this skill dir). It explains the three shape-tracking mechanisms (shape-aware stubs, DSL functions, special handlers) and how to check each one. You need this context to make the Gate 1 audit meaningful — knowing whether an op exists in a stub is not the same as knowing whether its shapes are tracked.

Do NOT read style_guide.md yet. It is comparison material for the verification phase. Reading it now biases you toward known patterns before you have empirically probed this model's shapes.

Gate 1: Audit ops

List every nn.Module subclass and torch/F. function called in the model. Check each against the shape-aware torch stubs (.pyi files in tensor-shapes/torch-stubs/) and any shape DSL declarations in those stubs. The stubs attach DSL shape functions with @uses_shape_dsl(ir_fn), and the IR functions live in tensor-shapes/torch-stubs/_shapes.pyi and are imported from stub files as torch._shapes because torch-stubs provides the torch package for type checking. Add any missing stubs or DSL functions BEFORE starting the port. This step should take minutes — you are scanning the stub file for the op and, only when a decorator points there, the corresponding IR function.

Do NOT delegate this audit to code search agents or use web search for this step. The tensor-shapes/torch-stubs/ package and its _shapes.pyi DSL file are exhaustive for torch shape support. For each op in your list, check whether it appears in the relevant stub file and whether it has a precise generic signature, Self/Tensor[*S] return, or @uses_shape_dsl(...) decorator that refines a bare declared return. Use targeted file reads or repo-approved search scoped to known files/directories, not broad recursive shell search. You need to confirm presence and spot missing attributes (e.g., bias on Conv2d), not memorize every signature.

Paste your audit table in your response before proceeding to Gate 2:

## Gate 1: Ops audit
| Op | Stub location | Shape DSL decorator / IR fn (or "no decorator") | Status |
|----|---------------|------------------------------|--------|
| nn.Conv2d | tensor-shapes/torch-stubs/nn/__init__.pyi — generic [InC,OutC,K,S,P,D] | no decorator (stub generic) | ✓ tracked |
| F.adaptive_avg_pool2d | tensor-shapes/torch-stubs/nn/functional.pyi — bare declared return | `@uses_shape_dsl(adaptive_pool_ir)`, defined in `_shapes.pyi` | ✓ tracked (DSL) |
| ...

Filling the "Shape DSL decorator / IR fn" column requires checking whether the stub declaration has a @uses_shape_dsl(...) decorator. If it does, confirm the named IR function exists in tensor-shapes/torch-stubs/_shapes.pyi. Write "no decorator" only after confirming the stub declaration has no decorator — do not leave this column blank or write "check DSL".

If any op is missing or returns bare, fix stubs/DSL BEFORE proceeding.

Gate 2: Inventory the original

Write the inventory as a comment block at the top of the port file, using this exact format:

# ## Inventory
# - [ ] ClassName.__init__ — Dims: param1, param2; int: param3
# - [ ] ClassName.forward
# - [ ] function_name — utility, no tensors
# ...

**Every class, function, and method in the original file must appear, and

…