Development Environment#

For routine Linux work, use the qualified molsyssuite@uibcdf_3.14 Conda environment and install participating local clones with python -m pip install --no-deps --editable PATH. Follow the suite’s development workspace contract for the canonical recipe, bootstrap and current eligibility. Check Python, editable import origins and python -m pip check; report existing environment conflicts separately from test results. The shared environment’s outstanding dependency conflicts are tracked in MolSysSuite #82.

For a standalone environment, use the repository’s Conda profiles below.

Typical setup#

# From the repository root; this profile selects Python 3.14.
conda env create --name pyunitwizard-dev --file devtools/conda-envs/development_env.yaml
conda activate pyunitwizard-dev
python -m pip install --no-deps --editable .
python --version
python -m pip check

Local validation baseline#

python -m pytest --receptor=llm tests
make -C docs html

The full supported matrix remains Python 3.11–3.14; a local Python 3.14 run does not qualify another interpreter, platform or installed release artifact.

The development, test and release-gate profiles pin public Ackredit 0.9.0 py_0 to exercise the optional Pint/unyt attribution pilot. Ackredit remains optional for normal package use. Fresh-process absence tests run with the provider blocked, and provisional function-observer tests skip when that capability is unavailable in the public release.

Run python devtools/benchmark_backend_attribution.py with PyUnitWizard installed in the environment to measure the warmed attribution cost. The tool retains all seven samples per case and reports whether function observation is available. Public 0.9.0 measures the portable backend paths; development providers exposing the observer also measure function attribution. Timing does not include imports, activation or first registration.

The prepared-provider path reuses immutable backend declaration plans only while software versions and bibliographic values still match. Nested metadata edits invalidate the plan; every completed credit continues to call the provider, including its registry conflict checks and independent captures. Public 0.9.0 retains the portable registration/tracking path. Compare paired normally installed producer wheels with identical scientific/provider inputs and retain raw samples; a source profile locates allocations but does not establish timing improvement. This optimization does not change the optional provider minimum or promote the provisional APIs.