topometry

This is the only recipe here that exists despite the package being on a channel. conda-forge carries exactly one TopoMetry build, 0.2.1.1, which predates the topo.sc.fit_adata and topo.sc.preprocess API that the eLife version of record (13:RP100361) describes. So there are two fixtures: topometry-environment keeps the stale channel build, and topometry-1.1-environment runs this one, which is what topometry-cell-cycle actually replicates against.

It builds from the PyPI sdist rather than a GitHub archive, and that is a correctness choice. Upstream’s committed setup.cfg still says 1.0.2 while topo/version.py says 1.1.0 — the version bump was applied at publish time and never committed — so a git build would produce a package that reports the wrong version to anyone pinning a commit for reproducibility. The sdist’s topo/ tree is byte-identical to master@f2653fa, verified by diff, and carries the correct version string. The recipe asserts that version three ways in its tests rather than trusting it.

The recipe also declares eight runtime dependencies upstream omits — scanpy, anndata, hnswlib, pacmap, scikit-misc, leidenalg, python-igraph, and adjusttext — every one of which the paper-facing single-cell workflow needs. topo.sc.preprocess defaults to the seurat_v3 HVG flavour, which wants scikit-misc’s loess. That is the packaging gap the topometry profile flags, fixed at the recipe layer rather than worked around downstream, with an upstream dependency PR prepared.

Bumping the conda-forge feedstock to 1.1.0 is the real fix: it would lift the fixture to L3 and retire the two-environment split entirely.