petls-pytorch

petls-pytorch is the open engine that unblocks everything petls cannot: same method family, Apache-2.0, pure Python, and therefore a noarch recipe that is eligible for staged-recipes. Publishing it promotes this fixture to L3 or L4; the L1 grade reflects only that the recipe is still in-repo. See petls-pytorch for the software profile.

The fixture builds from our fork jmchilton/petls-pytorch @ v2 rather than the PyPI 1.0.2 sdist. v2 carries three fixes over upstream, and one of them is load-bearing for this corpus: simplex-tree boundary extraction now indexes every simplex, so isolated vertices no longer crash Complex(simplex_tree=) and Rips(distances=). That crash blocked the bipartite interface complexes the TopoDockQ featurizer needs, where an atom with no cross-interface neighbour inside the cutoff is routine rather than exotic. The fork also demoted tadasets from a runtime dependency to a benchmark extra. Upstream pull requests are pending.

Recipe and linux-64 lock are verified green.