v0.4.3

Highlights

• Added differential CI against pinned MetPy 1.7.1 across Python 3.10 through 3.13 in metrust-only, metpy, and metpy-xarray modes.
• Added a first-class calc delegation ledger with deterministic tests for delegated and fallback behavior.
• Added cookbook-style replay coverage for sounding, grid diagnostics, and xarray-heavy workflows.
• Refreshed GitHub Actions workflows to Node 24-safe action versions.

Fixes

• Fixed the Python benchmark storm-relative helicity harness.
• Fixed geospatial_laplacian local fallback behavior.
• Fixed smooth_window xarray unit handling.
• Fixed MetPy signature hook recursion when metpy.calc imports after metrust.calc.

metrust 0.4.2

metrust 0.4.2 pushes the metpy.calc drop-in work further from wrapper compatibility into broad runtime parity.

Highlights:

• Reach zero public metpy.calc signature mismatches against MetPy.
• Expand runtime parity coverage across the remaining 91 reviewed calculation paths, including sounding/profile diagnostics, thermodynamic layer helpers, interpolation/dataset wrappers, wind-profile utilities, and extra kinematics cases.
• Fix parity-sensitive runtime behavior in parcel_profile, parcel_profile_with_lcl, isentropic_interpolation, mixed_layer_cape_cin, reduce_point_density, get_perturbation, geospatial_gradient, potential_vorticity_baroclinic, and related xarray/lat-lon inference paths.
• Keep the Rust backend as the default calculation path while allowing a small parity-sensitive subset to delegate to MetPy when it is installed.
• Update README and docs to reflect public signature parity and the current compatibility model accurately.

Verification:

• python -m pytest tests/test_metpy_dropin_compat.py tests/test_python_compat.py tests/test_signature_parity.py tests/test_runtime_parity.py tests/test_runtime_parity_interp_dataset.py tests/test_runtime_parity_kinematics_extra.py tests/test_runtime_parity_remaining.py tests/test_runtime_parity_sounding_core.py tests/test_runtime_parity_thermo_layers.py tests/test_runtime_parity_utils_misc.py tests/test_runtime_parity_wind_profiles.py tests/test_gpu_backend.py -q -> 174 passed
• cargo check -q
• local wheel build: metrust-0.4.2-cp313-cp313-win_amd64.whl

metrust 0.4.1

metrust 0.4.1 fixes the MetPy compatibility regressions identified in the review and narrows the top-level compatibility wording to match the current guarantees.

Highlights:

• Fix plain NumPy relative-humidity fraction handling in RH-derived thermo functions.
• Restore common MetPy keyword forms for wind_direction(convention=...), precipitable_water(bottom=..., top=...), and frontogenesis(potential_temperature=...).
• Support interior moist_lapse(reference_pressure=...) paths.
• Fix variable-spacing derivative, deformation, and frontogenesis parity on irregular grids.
• Add focused MetPy compatibility regressions for RH normalization, keyword forms, interior moist-lapse references, and irregular-grid kinematics.
• Update README/docs/package metadata to say MetPy-compatible calculation layer instead of claiming universal byte-for-byte wrapper parity.

Verification:

• python -m pytest tests/test_metpy_dropin_compat.py tests/test_python_compat.py tests/test_gpu_backend.py -q -> 30 passed
• local wheel build: metrust-0.4.1-cp313-cp313-win_amd64.whl

v0.4.0 — Optional GPU Backend

Optional GPU Backend (metrust[gpu])

metrust now supports an opt-in CUDA backend via met-cu. The default Rust/CPU backend is unchanged — existing code works exactly as before.

New API

import metrust.calc as mcalc

mcalc.set_backend("gpu")          # switch to CUDA
mcalc.set_backend("cpu")          # back to Rust

with mcalc.use_backend("gpu"):    # scoped switch
    theta = mcalc.potential_temperature(p, t)

Install with: pip install "metrust[gpu]"

11 GPU-Eligible Functions

• Scalar thermo: potential_temperature, equivalent_potential_temperature, dewpoint
• Grid kinematics: vorticity, frontogenesis, q_vector
• Grid composites: compute_cape_cin, compute_srh, compute_shear, compute_pw, composite_reflectivity_from_hydrometeors

All other functions automatically stay on the Rust CPU path.

Benchmark: Real HRRR Data (40 levels, 1059×1799, RTX 5090)

Category	MetPy	Rust	CUDA	CUDA/Rust
Scalar thermo (×10)	785 ms	136 ms	26 ms	5.3x
Grid kinematics (×5)	1.48 s	921 ms	32 ms	28x
Grid composites (×5)	—	3.72 s	1.32 s	2.8x
1D sounding (×6)	22 ms	0.67 ms	—	—
GPU-eligible total	—	4.50 s	1.37 s	3.3x

Highlights: frontogenesis 28x, q_vector 29x, vorticity 10x, compute_cape_cin 4.4x faster on GPU vs Rust.

Full Changelog

• Add get_backend(), set_backend(), use_backend() API
• Add metrust[gpu] optional dependency on met-cu>=0.2.1
• Add 3-column benchmark script (tests/benchmark_gpu.py) with real HRRR data
• Add 6 GPU backend unit tests