Skip to content

PROGRESS.md

Latest commit

 

History

History
164 lines (140 loc) · 7.69 KB

PROGRESS.md

File metadata and controls

164 lines (140 loc) · 7.69 KB

0.1.1 overnight progress notes

Branch: release/0.1.1 (off main). Spec: AUDIT_REPORT.md on branch audit/strategic-plan-2026-04. Date: 2026-04-07.

What I did

All work is on release/0.1.1. One logical fix per commit, test-first within each. Run git log main..release/0.1.1 --oneline for the order.

Bugs fixed (audit B1-B11)

  • B1 — CHILmesh.write_to_fort14 recursed into itself with the wrong arity. Now delegates to module-level write_fort14. The audit said "infinite recursion"; in practice it raises TypeError before recursing because the call passes 4 args to a 2-3-arg method. Same root cause, same fix. Regression: parametrized roundtrip identity test in tests/test_fort14_roundtrip.py.
  • B3 — _elem_type no longer treats vertex id 0 as a sentinel for "missing 4th vertex". Replaced the per-element loop with a vectorized "any pair of vertices in the row is equal" mask. Regression: synthetic quad with vertex 0 in slot 3 (tests/test_elem_type.py).
  • B4 — _ensure_ccw_orientation now classifies each CW row before flipping. Padded triangles get a triangle-only flip with the pad re-applied to slot 3; real quads get the original [0,3,2,1] permutation. Regression: test_ccw_mixed_element_flips_padded_cw_triangle_correctly, which fails on the previous code.
  • B5 — interior_angles quad branch now uses the same +1e-12 epsilon guard as the triangle branch and takes np.real of the result. Regression: zero-length-edge quad (test_quad_angles_no_nan_on_degenerate_quad) plus angle-sum invariants over every fixture.
  • B6 — Added src/chilmesh/utils/__init__.py. Regression: tests/test_package_imports.py::test_utils_subpackage_importable.
  • B7 — Stripped requirements.txt from a 60-line MADMESHR ML freeze to the three runtime dependencies declared in pyproject.toml.
  • B8 — README now uses pip install -e . from a checkout instead of the broken pip install requirements.txt.
  • B9 — README quickstart rewritten around chilmesh.examples.annulus(), removing the /kaggle/working/... path that no pip user could run.
  • B10 — pyproject.toml gained keywords, classifiers, [project.urls], requires-python>=3.10, and a [project.optional-dependencies] dev group.
  • B11 — Test count went from 6 to 58, and now parametrizes over all four shipped fixtures instead of just the annulus.

B2 — investigated, no fix needed

The audit reported that the quad signed_area shoelace "uses 4 cross terms instead of 8 → wrong by factor of ~2". On inspection the existing 4-row formula expands to the correct 8 cross terms:

0.5 * (x0(y1-y3) + x1(y2-y0) + x2(y3-y1) + x3(y0-y2))
  == 0.5 * sum_i (x_i*y_{i+1} - x_{i+1}*y_i)

Verified analytically, by hand on a unit square (area = 1.0) and a random parallelogram (area = base × height), and by the new test_signed_area_quad_shoelace_matches_numpy test which cross-checks against an independent shoelace on randomly-generated convex quads. Recommend leaving the existing implementation alone. The tests are kept as a hardening guard.

If you disagree, the next move is to re-derive the formula explicitly against MATLAB .m:1900-1908 (which I don't have access to in this environment).

Discovered out-of-audit bugs

  • read_from_fort14 could not parse legacy fort.14 files where the node indices were written as floats like 1.000000 (this affected structuredMesh1.14, one of the four shipped fixtures). One-line fix: int(float(...)). Without this, chilmesh.examples.structured() would have raised, breaking the parametrized test gate.

Infrastructure / packaging

  • Moved all four .fort.14 fixtures from doc/domains/fort_14/ to src/chilmesh/data/ and ship them via [tool.setuptools.package-data]. The directory now also has an __init__.py so it works as both a regular package and an importlib.resources source.

  • Added chilmesh.examples with annulus(), donut(), block_o(), structured() factories, plus a fixture_path() helper. Imported from the package root so chilmesh.examples.annulus() works.

  • Bumped version to 0.1.1 and requires-python to >=3.10.

  • Added tests/conftest.py with parametrized mesh and fresh_mesh fixtures and a session-level memoization layer that patches chilmesh.examples so each fixture is loaded at most once per test run. Block_O alone takes ~30s on first load (the O(n²) _build_elem2edge is the culprit; deferred to 0.1.2 per audit).

  • CI workflow staged at ci/ci.yml.pending (Py 3.10/3.11/3.12 × ubuntu/macos): installs editable, runs pytest, builds sdist+wheel, runs twine check, smoke-tests chilmesh.examples.annulus() from the installed wheel in a fresh venv. No coverage gate yet (deferred to 0.1.2 per audit). No release.yml yet — that needs your morning approval. Action required: the GitHub PAT used by this session lacks the workflow scope, so I could not push the file directly into .github/workflows/. Move it manually:

    mkdir -p .github/workflows
git mv ci/ci.yml.pending .github/workflows/ci.yml
git commit -m "ci: enable GitHub Actions workflow"
git push

    …or rerun this push from a session whose PAT has the workflow scope.

  • Added CHANGELOG.md.

  • Added a minimal .gitignore for build artefacts.

What I did NOT do (deliberately, requires your approval)

  • Push to main. All commits live on release/0.1.1.
  • Open a PR. You said you wanted to read the diff and open it yourself.
  • Tag v0.1.1.
  • Publish to PyPI.
  • Yank chilmesh==0.1.0 from PyPI (audit recommends yes; this is irreversible-with-effort and needs your call).
  • Configure PyPI Trusted Publishing (audit Open Question #1).
  • Add .github/workflows/release.yml — depends on the trusted publishing decision above.

Open question on _mesh_layers IE divergence (audit Q3)

I did not spend time on this. Read the audit guidance and the existing Python _mesh_layers produces a disjoint cover with monotone- shrinking layer sizes on the convex annulus and structured fixtures (verified by test_layers_disjoint_cover). The implementation it picks is "expand outward from boundary, peel two layers at a time (OE then IE)". The existing tests pin its current behaviour as an invariant rather than a value, which is the audit's preferred resolution. TODO for you: decide whether you want this annotated in a comment block in _mesh_layers as the chosen definition; I left the existing implementation untouched.

Test count

Audit asked for 8 new tests minimum. I shipped:

File Tests Notes
tests/test_package_imports.py 4 B6 + examples module
tests/test_fort14_roundtrip.py 4 B1 over 4 fixtures
tests/test_signed_area.py 7 B2 invariant + synthetic
tests/test_elem_type.py 2 B3 + mixed-element triangle pad
tests/test_ccw.py 8 B4 + idempotency over 4 fixtures
tests/test_interior_angles.py 6 B5 + sum invariants over fixtures
tests/test_invariants.py 5×4=20 layer cover, connectivity, etc.
tests/test_layers_annulus.py 6 pre-existing, repointed at examples
Total 57

How to verify locally

git checkout release/0.1.1
pip install -e ".[dev]"
pytest -v

The full run including Block_O takes ~3 minutes on my box; without Block_O it is under 10 seconds (pytest -k 'not block_o').