Changelog¶

All notable changes to this project are documented here. The format is based on Keep a Changelog, and this project adheres to Semantic Versioning.

[Unreleased]¶

First PyPI release of the fork.

Repaired the breakage left behind by the spectral/ann/dists simplification: restored imports and call sites (pairwise_distances now delegated to scikit-learn, spectral initialization via EigenDecomposition/LE), broke the kernels <-> spectral.eigen circular import, removed the dead graph_kernel state, and made the UoM path populate P_Z_/P_msZ_.
kNN graphs no longer silently drop genuine zero-distance neighbors (duplicate points); off-diagonal zeros are clamped to a tiny positive float so every row keeps exactly k edges (both sklearn and HNSWlib backends).

Deleted topo.base.dists; the Euclidean gradient is delegated to umap.distances.euclidean_grad.
Runtime dependencies slimmed: Jupyter packages moved to the dev group; hnswlib and pacmap are optional extras only ([ann], [layouts]), with corrected install hints.

API reference now covers the standalone building blocks (compute_kernel, eigendecompose, LE, graph_laplacian, diffusion_operator, automated_scaffold_sizing, projection diagnostics).
cknn_graph gained a full NumPy-style docstring with the Berry & Sauer reference. Example notebooks updated to the current API and re-synced with their jupytext sources.

Added integration tests composing the standalone estimators outside TopOGraph, plus a duplicate-point kNN regression test. Suite: 213 tests, ruff/mypy/pyright clean.

Renamed the distribution to topometry-nosc (independently maintained fork) to avoid colliding with the upstream topometry project on PyPI. The import package is still topo. Reset the version to 0.1.0.
Added the py.typed marker so the declared Typing :: Typed support is honoured by downstream type checkers (PEP 561).
Set fork authorship/maintainership, repointed project URLs to the fork (with an "Original project" link to upstream), and fork-ified CITATION.cff to cite this fork while referencing the original TopOMetry software and paper.
Migrated the build backend to Hatchling with dynamic versioning sourced from src/topo/version.py.
Adopted a src-layout (src/topo/) so tests run against the installed package.
Completed project metadata: real description, authors/maintainers, SPDX license, classifiers, keywords and project URLs.
Slimmed the core dependencies to numpy/scipy/scikit-learn/numba/tqdm. Moved plotting, dataframe and third-party layout/ANN dependencies into optional extras: plot, pandas, ann, layouts, notebooks, and all.

Documented the entire public API to a single NumPy-convention standard, added type hints across public signatures, and enabled ruff D (pydocstyle) lint enforcement to keep docstrings consistent.

Split the monolithic TopOGraph (≈1370 → ≈750 lines) into focused mixins under topo._pipeline (GraphBuildMixin, EigenBuildMixin, LayoutBuildMixin), mirroring the existing UoMMixin pattern. Public API, attributes and method-resolution order are unchanged.
Centralised optional-dependency and ANN-backend detection in topo._optional (single source of truth replacing duplicated _have_* flags). Missing optional dependencies now raise a clear pip install topometry-nosc[<extra>] hint instead of a raw traceback.
Replaced print() calls in library code with the topo logger hierarchy; verbosity now maps onto log levels via topo.configure_logging.
IntrinsicDim.transform is now a clean no-op for scikit-learn compatibility (previously printed a placeholder message); invalid random_state values now raise TypeError.

CI rewritten to use uv with a Python 3.10–3.13 matrix, plus ruff lint, ruff format check and a minimal-core import smoke test.
Added a tag-triggered release workflow (builds sdist + wheel; PyPI publishing stubbed and ready to enable).
Added Makefile, CONTRIBUTING.md, ruff/mypy/pytest/coverage configuration.

Single-cell / scanpy / AnnData wrappers — TopoMetry is now a standalone geometry toolkit.

Core API unchanged: TopOGraph, spectral scaffolds, graph operators, layouts, metrics and plotting.