Development Environment

Use conda environments from devtools/conda-envs.

Recommended environments

• development_env.yaml: core development tasks.

• test_env.yaml: test execution.

• docs_env.yaml: Sphinx and notebook documentation.

• release_gates_env.yaml: release-gate parity checks.

Typical setup

cd devtools/conda-envs
python create_conda_env.py -n pyunitwizard-dev -p 3.14 development_env.yaml
conda activate pyunitwizard-dev

Local validation baseline

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

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.