Pyrefly JAX shape stubs

This package is a PEP 561 stub-only distribution. It installs the jax-stubs stub package so Pyrefly can discover shape-aware stubs for the runtime jax package without replacing or shadowing JAX itself. It is versioned in lockstep with Pyrefly and depends on the matching pyrefly-shape-extensions package. Installing it also installs jax, which provides upstream modules used by these partial stubs. The jax dependency is unpinned, so projects can choose their own JAX version.

Shape-typed fixture stubs for a subset of JAX. This is a starting point for working with the JAX core team rather than a complete model: it covers array creation, broadcasting arithmetic, matmul, reshape, transpose, the axis/keepdims reductions, and the elementwise activations in jax.nn. Dimensions are modeled and dtypes are not, so adding dtypes later means modeling JAX's own defaults (float32 and int32) rather than copying the NumPy stubs.

Shape rules use the type-level DSL, @type_shape_dsl_function, exclusively. A rule the DSL cannot express yet returns a gradual shape instead. Where a rule is imprecise for that reason, the stub says so at the definition, along with whether a fix is expected.

TENSOR_SHAPES_CONTRIBUTING.md at the repository root covers the workflow, and tensor-shapes/run_tests.py runs the tests. Pyrefly checks the stubs themselves and every test/test_*.py, and those same test files then run against real JAX, so a stub that is self-consistent but wrong still fails.

Before submitting a change, format this package with Ruff from the repository root:

uv tool run --from ruff==0.16.5 ruff format tensor-shapes/jax-pyrefly-stubs

Anything not listed above is simply absent rather than modeled loosely, so it is reported as a missing attribute rather than inferred gradually.