petls-pytorch

petls-pytorch is the shippable answer to what petls-recipe cannot be: an independent PyTorch reimplementation of the same persistent-Laplacian engine, Apache-2.0, pure Python, and therefore noarch and eligible for staged-recipes. Publishing it lifts petls-pytorch-environment from L1 to L3 or L4; the L1 today reflects only that the build lives here.

It builds from our fork at v2, not the PyPI 1.0.2 sdist, and that choice is load-bearing rather than incidental. The fork carries three changes over 1.0.2: simplex-tree boundary extraction now indexes every simplex, so isolated vertices no longer break Complex(simplex_tree=), which the bipartite interface complexes in open-topodockq-featurizer-recipe require; tadasets moved from a runtime dependency to a benchmark extra, which is what makes the install closure solvable at all; and float64 property tests were added. The version stays 1.0.2 because upstream’s is static, so the build number is what distinguishes this package from a stock one.

The dependency demotion has since been merged and released upstream; the isolated-vertex fix was independently resolved by upstream’s weighted-alpha refactor. What remains fork-only is the test coverage. A future recipe that could point at an upstream release rather than a fork archive is the right end state, and is what to check before opening the staged-recipes PR.