triangular

Two-dimensional Delaunay triangulation and convex hull for Python, implemented in Rust. The API is compatible with the triangle wrapper's delaunay and convex_hull functions.

import numpy as np
import triangular

pts = np.array([[0, 0], [0, 1], [0.5, 0.5], [1, 1], [1, 0]], dtype=float)

triangular.delaunay(pts)      # -> int32 (M, 3): 0-based vertex-index triples
triangular.convex_hull(pts)   # -> int32 (K, 2): 0-based hull edge vertex pairs

These correspond to triangle.delaunay (Triangle options Qz) and triangle.convex_hull (options Qzc). Inputs are validated in Python: anything coercible to a float64 (N, 2) array is accepted; bad shapes, non-finite values, or fewer than three vertices raise ValueError. Degenerate inputs with no triangulation or hull (all points collinear, all points coincident) return an empty array of the correct shape. CONTRACT.md documents the exact behavior, including the few places it intentionally differs from triangle.

Layout

python/triangular/   Python package: input coercion, validation, error contract.
src/                 Rust geometry core, built with PyO3.
corpus/              Validation corpus recorded from the triangle library.
tests/               pytest suite: equivalence, invariants, degenerate cases, benchmarks.
CONTRACT.md          Spec for the native core.

Development

uv sync          # installs dev dependencies and builds the Rust extension
uv run pytest

uv sync builds the native module via maturin; after changing the Rust code, rebuild with uv sync --reinstall-package triangular.

The suite can also run against the real triangle library, to check that the corpus and invariant tests are themselves correct:

TRIANGULAR_IMPL=triangle uv run pytest

Tests marked improvement (where triangular's degenerate-input behavior intentionally differs from triangle's) are skipped in that mode.

Benchmarks

uv run pytest -m benchmark

Compares against the reference timings recorded in corpus/manifest.json. The suite reads a few environment variables:

Variable	Default	Meaning
TRIANGULAR_IMPL	triangular	Which import package the suite exercises.
TRIANGULAR_BENCH_TOLERANCE	1.5	Max allowed triangular / triangle time ratio.
TRIANGULAR_BENCH_MAX_N	100000	Skip benchmark cases larger than this.
TRIANGULAR_FULL_CIRCLE_MAX	1500	Brute-force empty-circle check size cap.

Linting

pre-commit runs ruff and pyright on Python and rustfmt and clippy on Rust:

uv run pre-commit install
uv run pre-commit run --all-files

Releasing

Versions live in git tags only — pyproject.toml stays pinned at the 0.0.0 placeholder and is never edited for a release. To ship:

gh release create v0.3.0 --generate-notes

Publishing the GitHub release triggers .github/workflows/release.yml, which stamps the version from the tag into pyproject.toml, builds wheels (linux x86_64/aarch64, macOS universal2, windows x64 — all abi3-py312, one wheel per platform) plus an sdist, smoke tests each wheel, and publishes to PyPI via trusted publishing. Tags must match vX.Y.Z (optionally aN/bN/rcN suffixed); anything else fails the stamp step before a build starts.