Skip to content

0.4.3 candidate decision record — 2026-10-16

This is a release gate, not an announcement that a version has been tagged or that a general forming solver has been validated. The initial moving-facet evidence below came from af7b52a3; final release evidence must instead bind to the integrated candidate's exact source and installed wheel.

Post-release outcome — 2026-10-17

PR #114 merged as f6f9ea431a5307bbfcc9bfb85da2bfa82ce51b64 and the v0.4.3 tag points to that commit. The exact integrated source passed the Linux FEniCSx/MPI and macOS installed-wheel checks; publish run 38944199683 completed successfully. PyPI serves the immutable Linux wheel with SHA-256 cea6e7f6bbb4e01df3129169c97b5976a6b7354eab4c30fd58a06e1424720fab and source archive with SHA-256 bd5f4f255634ccaa2e4389d94014db549ffb23d22520d1eb9e16b29a3654b161. Those hashes match the run artifacts. The repaired archive identity check passed despite normal directory records in the Linux wheel. This closes the 0.4.3 publication gate, not the scientific exclusions below; the v0.4.2 GitHub-only release was not retroactively published to PyPI.

Proposed bounded claim

One serial, three-dimensional finite-strain logarithmic-J2 displacement Step can compose with one frictionless penalty rigid tool through the ordinary model.step(...) workflow. The tool may be an analytical plane or a triangulated surface whose active contact remains on a planar facet or a coplanar interior seam. Its prescribed rigid motion uses the Step's normalized load path. The same accepted increment owns the J2 material state, tool pose, reaction, contact potential, force--moment work, result, and restart identity.

This does not admit curved or creased active facet transitions, friction, multiple pairs, contact/strong-BC overlap, mixed displacement--pressure, periodic MPC plus contact, or distributed implicit contact. Provider discovery and Step lowering must reject unsupported combinations before residual assembly; no boundary model may be silently dropped. The composed result's computed trust status is intentional. A package version cannot promote it to industrial forming verification.

Evidence already obtained on the candidate branch

  • Analytical plane and fixed/moving coplanar facets agree on the bounded J2 fixtures for displacement, committed state, contact potential, generalized force and accepted tool work. The new moving-facet dual matches the fixed-state potential derivative at two z-axis rotation magnitudes. A separate three-axis proportional rotation check also matches the analytical plane and its force--work derivative; its regression is included with the pending release candidate. A separate no-traction run confirms that prescribed tool motion alone drives the body, rather than relying on an unrelated applied force. Strong/affine boundary rejection, failed-attempt rollback, serial restart, and the two-rank fail-closed path have targeted regressions. See the fixed-facet and moving-facet records for exact scope and tolerances.
  • A 0.2-rad fixture has 3/6/12-increment and 1--4-cell integral-output sensitivity records. Consecutive tool-work and penalty-potential differences shrink in those scans; spatial independence is not established, the energy residual is not monotone, and the penalty scan changes the physical result.
  • A native macOS arm64 CPython 3.11 wheel and sdist were built from af7b52a3. The wheel SHA-256 is 7e03d5b5c419afde39f32ef3a418858496aa3a26c9665ca4c2d6cc60bbb44b9d; the sdist SHA-256 is 717a0e4990f5b38b0317d0ae169335784023c7acb52303b942ba23c904a1dac2. The isolated installed-wheel release_gate.py --dist ... --smoke --require-platform macos passed. From that installed wheel outside the checkout, examples/finite_j2_faceted_contact.py --tool-travel=-0.02 --tool-rotation=0.2 completed, wrote XDMF/HDF5 and a result manifest, and reported contact potential 1.04376 and tool work 2.00687 in its dimensionless fixture. These are smoke observables, not calibrated forming predictions.

Still required before tagging

The previous v0.4.2 GitHub release did not reach PyPI: its Linux repaired wheel contained 37 ordinary ZIP directory records, which the installed-wheel identity gate incorrectly compared as runtime files. The 0.4.3 candidate now filters archive directory records and retains the actual file-hash comparison; a regression exercises both ZIP and source archives. The exact tagged Linux wheel must still pass the publication workflow's installed-use gate before PyPI upload. A GitHub release page alone is not publication evidence.

  1. The analytical-plane, fixed-facet and moving-facet PRs have merged. Required remote checks on the exact 0.4.3 release PR head, including installed FEniCSx, must finish successfully before that PR merges.
  2. Freeze the integrated source and run candidate-bound installed-wheel checks after the release metadata merge; do not reuse a branch-wheel hash as evidence for a changed source tree.
  3. The changelog, version, citation and scoped release notes move together. Tagging triggers the publishing workflow; verify that exact tagged wheel, then confirm actual PyPI publication rather than assuming a GitHub tag did it.
  4. Preserve the known limits in capability discovery and user documentation. Windows/WSL or a nonplanar die needs separate installation and scientific evidence; a Mac wheel smoke cannot establish either.

Next numerical gate, not a 0.4.3 release shortcut

For a real curved triangulated tool, add smooth-surface or reviewed edge/vertex branch derivatives, then compare with an independent structure reference while refining tool facets, solid mesh, increments and penalty. Abaqus/Standard's analytical-rigid-surface guidance also flags discontinuous normals and tangents as a convergence difficulty; this supports the present fail-closed branch policy, not a claim that our piecewise-planar law is equivalent to its general-contact implementation. For periodic MPC, assemble the total material-plus-contact residual and tangent before applying T^T R and T^T K T on every Newton and line-search trial, with joint state/work/restart evidence. Passing a contact parameter to the current material-only affine route would be incorrect.