Repository navigation
Releases: audunarn/ANYsolver
Release list
ANYsolver 0.4.7 — accepted performance and runtime fixes
ANYsolver 0.4.7 packages the accepted static, spectral, and nonlinear performance work since 0.4.6, plus runtime failure-status corrections.
- Representative complete static analysis improved 3.98× and retained load-only analysis 19.85× on the declared workload.
- Retained medium-panel spectral analysis improved 11.52% in the seven-pair gate, with identical recorded modes and residuals.
- The representative nonlinear gate improved its four non-easy cases by a 12.44% median with passing physical comparisons. The follower-shell case was 0.96% slower.
- Force-control reaction reuse, fixed dead-load preparation, trial-tangent promotion checks, and solver-work diagnostics are included.
- The generated-geometry runtime facade preserves constraints and nonlinear failure details for locally edited models.
The Armijo line search is optional and remains experimental. Its separate convergence promotion decision was NO-GO; no convergence benefit is claimed. Element equations, physical tolerances, defaults, and coordinated dependency bounds are unchanged.
The full PR and post-merge CI, installed-package gate, and TestPyPI rehearsal bind this release to the reviewed source. See PR #58, release details, and the TestPyPI staged package.
ANYsolver 0.4.6
ANYsolver 0.4.6: bounded ecosystem compatibility release
ANYsolver 0.4.6 widens the supported mesher line to ANYmesher>=0.4,<0.6
and raises the file-I/O floor to ANYfileio>=0.3.2,<0.4. The bounded gate
uses production ANYmaterial 0.2.0, ANYgeometry 0.4.3, ANYmesher 0.5.0, and
ANYfileio 0.3.2 artifacts.
The release changes dependency metadata and compatibility evidence only.
Solver mechanics, B3-GE explicit opt-in behavior, the legacy B3 default, and
qualified Q4/S3 mechanics and defaults are unchanged.
The gate validates the immutable B3-GE G7 evidence, exact installed package
identities, source and wheel runtime equality, dependency resolution, neutral
meshing and file-I/O public contracts, licensing, selector boundaries, and
the trimmed source archive. It does not rerun scientific qualification.
Terminal on success:
PROVISIONAL_GO_ANYSOLVER_0_4_6_BOUNDED_COMPATIBILITY_RELEASE
This terminal authorizes a separate publication decision for the exact 0.4.6
artifacts; it does not publish them automatically.
ANYsolver 0.4.5
Bounded maintenance release for the accepted B3-GE production opt-in. Legacy B3 remains the default; qualified Q4 and S3 mechanics and defaults are unchanged. The dependency-complete installed-wheel gate, licensing checks, exact G7 evidence binding, and trimmed source distribution all pass. This successor preserves the failed v0.4.4 GitHub publication attempt; no v0.4.4 artifacts reached PyPI. Compatibility remains ANYmesher >=0.4,<0.5 because published ANYfileIO 0.3.1 rejects ANYmesher 0.5 semantic operations.
ANYsolver 0.4.4
Bounded maintenance release for the accepted beam qualification work. B3-GE is available only through the explicit b3-ge selector; legacy B3 remains the default. Qualified Q4 and S3 mechanics and defaults are unchanged. The release adds a deterministic installed-artifact gate and trims obsolete internal development material from source distributions. Compatibility remains ANYmesher >=0.4,<0.5 because published ANYfileIO 0.3.1 rejects ANYmesher 0.5 semantic operations.
ANYsolver 0.4.3 — scoped native beam workflows
Current accepted functionality
- Explicit
ge-beam3straight mixed beam andge-beam3-nativemodel-owned straight/regular-curved workflows. - Declared generalized and physical-fibre history, recovery and authenticated restart; admitted reference/current-rest modal and frozen-current buckling workflows.
- Reviewed objective eccentric generalized-beam connection to one owned elastic qualified Q4 or S3 V2D.
- Existing B2/B3, qualified Q4/S3 mechanics and all defaults are unchanged.
This is scoped opt-in delivery, not qualification of generic mixed FEModel integration, plastic/fibre-shell coupling, arbitrary sections, finite-velocity rotational dynamics or stable postbuckling from rest.
Verification
PR #43 passed all 24 normal CI checks. The release runtime differs from the independently accepted delivery only in the version assignment. All 317 runtime files are checked against the accepted manifest before trusted publication. The isolated release wheel reproduced all 20 accepted scientific records byte-for-byte. Historical qualification records and incidents remain immutable and are preserved in the merged ancestry.
API and limits | Release verification | CI
Publication completed through the existing trusted-publisher workflow. The wheel downloaded directly from PyPI passed isolated public-workflow replay: all 20 scientific files are byte-identical to the accepted evidence; process tree empty, about 16.89 seconds. No dependency upgrade for the concurrent ANYmesher 0.5 candidate is included.
Published artifact verification
Main and tag commit: 74703a3202251edc0beafb21cd31f52c9304ceb8.
Trusted publisher: https://github.com/audunarn/ANYsolver/actions/runs/34404094397
PyPI: https://pypi.org/project/ANYsolver/0.4.3/
| Artifact | Bytes | SHA-256 |
|---|---|---|
| anysolver-0.4.3-py3-none-any.whl | 1,774,463 | e7b453bd070f82c5615fa3d1879530eeaf60b7f23b6875a3966a94855bd2cd05 |
| anysolver-0.4.3.tar.gz | 4,585,022 | 9a43d18c12d3fcc8d38c3f7c6f1d73288c996e0737774942187297e60981663d |
These exact files match PyPI and are attached below with the runtime-bridge receipt. This receipt is byte-identical to the verification record produced before trusted publication. The release does not rewrite or broaden the historical qualification.
ANYsolver 0.4.2
Canonicalize solver-owned compact qualified-Q4 plastic state at the accepted Newton finalization boundary so committed final states replay exactly after valid plastic increments. Public sealing remains fail-closed. Qualified Q4/S3 mechanics, coefficients, recovery, formulation identities, tolerances, and defaults are unchanged.\n\nValidated by all 24 CI jobs plus exact ANYfem capacity source and installed-wheel regressions.
ANYsolver 0.4.1
Repairs qualified E4-PL Q4 accepted-state replay after valid plastic increments and aligns the coordinated ANYmaterial 0.2, ANYmesher 0.4, and ANYfileio 0.3 dependency graph. Qualified Q4 and qualified S3 defaults remain unchanged.
ANYsolver 0.4.0
ANYsolver 0.4.0 is the first source release under the Mozilla Public License 2.0. Earlier releases retain their original licence terms.
This release includes the qualified E4-PL Q4 default and the qualified E4-PL S3 V2D companion for current-policy three-node shells, while retaining explicit legacy rollback and migration behavior. It also adds canonical dependency-licence notices and automated licensing/artifact compliance gates.
Known boundary: qualified S3 admission fails closed at automatically imprinted four-way shell junctions until junction-aware meshing support is available; there is no silent legacy fallback.
See CHANGELOG.md, LICENSING.md, and THIRD_PARTY_NOTICES.md for details.
ANYsolver 0.3.0
ANYsolver 0.3.0 publishes the extracted solver architecture and the accumulated Unreleased changes documented in CHANGELOG.md.
Highlights include generalized shell and beam section contracts, orthotropic and Hill-48 material support, performance_2 solver acceleration, bounded analysis-session reuse, improved recovery, and preserved fail-closed S4 qualification evidence.
Production restrictions and the legacy ShellElement default remain unchanged.