Tests, builds and fast feedback

Routine development uses focused local checks and never triggers or waits for CI. Full validation is required for release candidates and remains available locally for a deliberate run.

Inventory

The pre-change snapshot at 6890700c6a32 (2026-09-26) has 172 test files and 2,506 test definitions. Local collection produced 3,989 parameterized cases: 3,980 selected and nine benchmark cases deselected. Optional dependencies affect collection/skips. The new CI-scope tests are additional to this baseline.

The routine smoke inventory now executes 79 cases in six files, down from 293 in the previous sample. A local run with rendering/optional imports blocked and external pytest plugins disabled passed in 6.51 seconds. Nine cell cases were also removed from the sample because one imports the VTK rendering stack. The previous MR test job took 172.59 seconds including setup. These are observed runs, not a controlled equivalent-host speedup test.

Kind	What is actually exercised
Unit/contract	Units, kinds, cells, bonds, configuration, archive metadata/schema, code and release metadata guards
Reader/API integration	Real temporary QVF/ZIP files, payloads/checksums, ownership/closing, corruption handling, conversion and CLI behavior
Rendering/application integration	VTK/PyVista objects and offscreen renders, controller state, browser/TUI behavior, geometry and volume transformations
External/optional integration	QC wavefunction/relocalization and live optimization; optional format/tool integrations; served-browser tests need a browser
Benchmarks	Nine cases excluded by the normal marker default; opt in with pytest -m benchmark
Build checks	Wheel/sdist reproducibility, clean installed-wheel smoke and dependency integrity; Sphinx site build
Static checks	Ruff, and full strict mypy in full-validation pipelines

This is a pure-Python package. CI installs prebuilt rendering dependencies; it normally does not compile a quantum-chemistry core or run a chemistry campaign. QC-dependent tests can skip here. A pinned subset runs with a real QC core in the producer’s release gate. Full viewer CI is not proof that every optional browser or QC-dependent route ran.

Pipeline matrix

Context	Ruff	Primary Python 3.13	Mypy + Python 3.12 + 3.14	Wheel/sdist + installed-wheel smoke
Development refs and MRs, including web/scheduled/API requests	No pipeline	No pipeline	No pipeline	No pipeline
Release candidate, any trigger	Yes	QVF conformance + full normal suite	Mypy + compatibility samples	Yes
Release branch	No repeated runtime gate	Docs publication build only	No	No

The complete behavior suite runs once on Python 3.13. Python 3.12 and 3.14 run the nine-file COMPATIBILITY_TESTS sample: QVF read/write/schema, API, CLI, renderer dispatch, controller calls, units and kinds. This removes two repeated full-suite executions while retaining supported-interpreter integration checks. Detailed feature regressions and optional integrations run only on 3.13; this sample does not claim full behavior parity on every interpreter. Expand it when a version-specific regression needs coverage.

At the 2026-10-25 snapshot, local collection selects 7,048 cases in the full suite and the compatibility sample passes 272 cases in 17.08 seconds on Python 3.14. This changes the three-interpreter test budget from approximately 21,144 to 7,592 case executions (64% fewer). Dependency installation and platform skips vary, so this is a case-count reduction, not a measured CI wall-clock speedup.

A further 2026-10-25 reduction removes 271 collected cases from the main suite:

• Single-command archive checks use seven representative error boundaries; table, validate and the API/reader paths retain every corruption kind. Diff still checks both argument orders and both output modes, but does not cross every low-level ZIP variant with every formatting option.

• SMILES successful-build assertions share one app instead of constructing four, and normalization assertions share the Gram/Cartesian checks.

• Normalization grids use 48 rather than 160 points per axis (97.3% fewer points). All angular momenta, components and tolerances remain; observed norm errors stay below 4e-15.

The three affected modules went from 731 passed / five skipped in 21.27 seconds to 463 passed / two skipped in 15.05 seconds locally. These runs share a machine but are not controlled CI benchmarks. The broad seeded fuzzers and targeted historical failures remain. The removed command/error combinations can no longer identify an interaction confined to that exact pair; the retained cases cover container, schema, critical-section, filesystem, binary-member and JSON-member failure boundaries.

The fast sample covers QVF open/write/schema, units, kinds and privacy. It omits the broad reader-ownership, API, bond/render/controller and CI-policy suites from routine runs. There is no automatic changed-test expansion or Git history lookup. Run affected tests locally or use --scope changed deliberately; in an MR environment that adds changed test files (and selects document contracts for prose-only diffs). Full validation still includes these tests.

Local smoke setup needs the core package and pytest. Candidate CI always installs the test/release extras, external QVF corpus and GL/Xvfb stack, and runs conformance, one full suite, two compatibility samples and packaging unconditionally. A test-scope variable cannot downgrade it to smoke. Docs build strictly on candidates and publish from release; development docs changes use a local check. Candidate pipelines disable automatic cancellation.

Run or inspect a lane

The QVF check uses the existing QVF_TAG clone from GitLab and compares vendored bytes against that release. CI installs the reference validator from that same clone’s python[validate] package before the test extras. For local checks, install it from your checkout of the pinned release first:

python -m pip install '/path/to/qvf/python[validate]'
python -m pip install -e '.[test,validation]'
python scripts/check_qvf_conformance.py --qvf-root /path/to/qvf --strict-negatives

The candidate lane always uses --strict-negatives. Positive and legacy archives must validate, open, decode vibration/GTO payloads and parse viewer hints. Every negative must produce the expected error or warning codes from its .expected.json; resource-cap fixtures use that file’s limits overrides. Warning-only fixtures must remain valid. Exceptions, wrong acceptance, missing codes and schema drift fail the strict check. No corpus entries are skipped.

A previous v0.4.0 corpus run passed 76/76 positive and legacy archives and 110/110 negative expectations, with zero gaps or unexpected exceptions. Those counts are historical evidence, not a result for the current v0.6.0 pin. Use the command above with the current QVF_TAG checkout to obtain evidence for that corpus. Without --strict-negatives, the script still reports missing findings as gaps for diagnosis. Format and resource rules come from qvf’s reference validator, not a second implementation in vibe-view. Neither lane edits the corpus or the vendored schema.

PYVISTA_OFF_SCREEN=True python scripts/run_ci_tests.py --scope fast
python scripts/run_ci_tests.py --scope fast --list-targets
PYVISTA_OFF_SCREEN=True python scripts/run_ci_tests.py --scope changed  # opt-in MR diff expansion
python scripts/run_ci_tests.py --scope full
PYVISTA_OFF_SCREEN=True python scripts/run_ci_tests.py --scope compatibility
python -m pytest tests/test_<area>.py -q
python -m pytest tests --collect-only -q

FAST_TESTS and PROSE_TESTS in scripts/run_ci_tests.py are the explicit sample. The full suite is retained: no bulk deletion of regression tests. GitLab web, scheduled and API requests create pipelines only on release refs; target a release-candidate/* branch to repeat full qualification. JUnit and the 30 slowest phases are published for the runtime jobs. Package artifacts are produced only by the full primary job.