Repository navigation
Releases
v1.7.0
v1.6.0 — Nonconvergence policy and verification loop
- Opt-in block-solver nonconvergence policy (#14): nonconvergence_policy="continue" commits maxiter-exhausted local solves, counts them in checkpointed progress and reports local_solves_converged; breakdowns always fail; default fail unchanged.
- Verification loop (#15): tools/verify.py (public API, solver golden hashes, README examples, cross-version checkpoint resume vs the base branch), REVIEW.md diff-triggered checklist, CI runs the base commit's verifier.
Scientific operators unchanged; source identical to 1.5.4, which winding-field-cli 1.6.0 pins.
Full changelog: v1.5.3...v1.6.0
v1.5.3 — Optional advisory observers
winding-field v1.5.3
Changes
- Add optional advisory transaction/sweep observers for clients that request progress diagnostics.
- Preserve observer delivery across interrupted/resumed closures and keep advisory clocks separate from existing stop-boundary checks.
- Release temporary reduction arrays promptly; include formatting, lint and observer typing corrections.
- Promote the already-merged observer-capable development version to stable package metadata
1.5.3. This release preparation changes no numerical module bytes.
Verification and limits
- The preceding
mainrevision passed remote CI: https://github.com/LimeGS/winding-field/actions/runs/44189936875. - Release metadata preparation passed 283 focused distribution, observer, physical-grid, store, runtime, solver and interface tests locally.
- Built wheel and sdist; checked metadata, packaged module bytes and retained numerical source hashes. CLI compatibility qualification is reported separately.
- No local full-suite rerun, GPU/real-volume qualification, PyPI publication or deployment is claimed.
- Existing runs/checkpoints and installed runtimes are not migrated or rewritten; cross-version resume compatibility is not promised.
v1.5.2 — Exact physical stencil boundaries
Fix a physical-grid boundary rejection: valid final-node coordinates could normalize one floating-point step above the node limit, causing constraint point outside valid support. Validate strict physical bounds first, then clip normalized coordinates for interpolation. True out-of-domain points remain invalid; no tolerance expansion, solver change or rescaling.
Critical verification: 97 core tests passed; 62 CLI stride/field tests passed against an isolated core wheel, including the previously failing 172.8 µm stride16 pipelines. Independent critical review found no actionable issues. Full suite and CI wait omitted at user request. No unrelated unmerged solver changes included.
PR: #9
v1.5.1 — Solver performance and CLI-ready numerical core
winding-field v1.5.1
The numerical core and CLI now share the v1.5.1 release version.
Changes since v1.2.0
- Consolidate the numerical library under the
winding-fielddistribution; CT workflows, mesh extraction and rendering are owned bywinding-field-cli. - Reuse bounded solver workspaces, PCG buffers and residual norms while keeping live targets, fixed values, input validation and the final true-residual check.
- Add prepared CPU consensus indices and opt-in spectral PCG for supported standalone CPU float64 problems. Block workflows remain Jacobi-only.
- Add bounded spatial/template/gather caches, remembered nonadmission decisions and optional coalesced status reporting without weakening durable checkpoints or resume identity.
- Reduce repeated durable-store decoding and retain result-preserving v2 optimizations and reproducible CPU benchmark harnesses.
- Integrate development dependency updates and compatibility regressions. The Lean pilot is explicitly excluded.
Verification
- Release commit:
96631f3cf8aad385054b58b48c91a05a6a053eb1. - GitHub Audit: Python 3.11/3.12 tests, offline benchmark, lint/format/type checks, build and wheel/sdist installation checks.
- Local core: 1,201 passed, 1 optional test skipped; 93.69% statement coverage. Installed wheel tests and isolated examples were also checked during integration.
- Downloadable wheel and sdist include
SHA256SUMS.txt.
Compatibility and scope
Python 3.11 and 3.12. Flat numerical module imports remain available; use the separately released CLI for workflow and surface commands. Cache budgets limit retained data, not total process memory. Optional spectral support is not available for arbitrary block masks.
Synthetic numerical/CPU checks do not certify physical sheet identity, full-scroll performance or new GPU qualification. Existing historical runs and releases are preserved.
Full changes: v1.2.0...v1.5.1
v1.2.0
Documentation and measured-results update for the experimental winding-field library.
Changes
- Replace historical README result tables with the latest Paris4 fine-band measurement at 19.2 µm: 792,595 of 2,154,184 frozen AFV v2 pairs covered (36.8%), with explicit coverage, calibration and checkpoint identities.
- Add the full-resolution six-surface Paris4 panel above the earlier CT/winding overlay, with consistent captions, attribution and data terms.
- Move the repository introduction before the images, simplify the measurement heading and remove stale comparisons and cross-references.
- Update the installable package version to
1.2.0.
Verification
- GitHub Audit passed for commit
360866c98d138334dfa842471c2d04c8026d5466: Linux Python 3.11/3.12 tests and offline benchmark quickstarts, formatting/lint, type checking, wheel contents, and wheel/sdist installation checks. - Local macOS / Python 3.11.13: 512 tests passed, 91.29%
srcstatement coverage. Ruff 0.17.0 formatting/lint and mypy 2.4.0 passed. - Wheel and sdist built, installed and smoke-tested in a dedicated virtual environment.
- README data rows checked against the saved evaluation result; image identity, ordering and documentation links checked.
- Existing tests pass with the updated README; no test changes or weakened checks were needed.
- Downloadable wheel and sdist are accompanied by
SHA256SUMS.txt.
Compatibility
The core solver API and numerical implementation are unchanged. The package targets Python 3.11 and 3.12. The wheel installs library modules; command-line tools and examples live in the repository checkout.
Scientific scope
The new table measures intrafibre consistency on a partial fine-band checkpoint, using finite stored fine values without coarse fallback for unavailable samples. It is a development measurement, not independent physical validation or a final convergence claim. The checkpoint-specific evaluation adapter and original execution artifacts remain in the research workspace.
The six-surface image is diagnostic and marked unaccepted; unavailable or unmapped regions remain explicit. CT, AFV inputs and derived illustrations retain their upstream terms.
Full changelog: v1.1.0...v1.2.0
v1.1.0
Dataset-independent winding-field construction, evaluation and surface workflows.
Changes
- Add explicit JSON dataset configuration for observation sources, AFVs, coordinate transforms, reconstruction grids and fixed calibration.
- Generalize construction, benchmark and surface tools, with provenance-bound caches, support-mask checks, explicit exterior anchors and solver-failure handling.
- Add portable benchmark cases and CT visualization geometry, including optional human references and explicit unavailable results.
- Add an offline construction demo, a dataset template and migration documentation. The demo uses a float32-appropriate convergence tolerance and checks the complete field against its analytic solution.
- Present the AFV v1/v2 fitting results side by side, with public-source credits to Qual and Bruniss and preserved measurement identities.
Compatibility and migration
Existing research CLI commands need migration: real-data workflows now require --config and explicit physical-coordinate declarations instead of implicit scan names, AFV paths or host directories. Some helper APIs also require explicit geometry. See the dataset configuration and migration guide.
The core field_solver.solve_field array API is unchanged. The installable package is winding-field-experiments==1.1.0, targeting Python 3.11 and 3.12. Command-line tools and examples live in the repository checkout; the wheel installs the library modules.
Verification
- GitHub Audit passed for release commit
f7821f567bc8a7ce11eda63b7229e56bbd29e2e9: Linux Python 3.11/3.12 tests and offline benchmark quickstarts, formatting/lint, type checking, wheel contents and wheel/sdist installation checks. - Local macOS / Python 3.11 and 3.12: 512 tests passed on each interpreter, 91.29%
srcstatement coverage. Ruff formatting/lint and mypy passed. - Wheel and sdist built, installed and smoke-tested in an isolated virtual environment.
- README comparison checked against frozen results: 11 paired bins, 66 drift values, coverage, human controls and AFV/field identities. Contributor credits checked against the original release posts.
- Downloadable wheel and sdist are accompanied by
SHA256SUMS.txt.
Scientific scope
This remains experimental research software. The AFV tables are development-consistency measurements pinned to their original execution source, not new real-data runs of the generalized adapters or independent physical validation. Real CT, AFVs, model weights and run artifacts remain external and retain their upstream terms.
Full changelog: v1.0.1...v1.1.0
v1.0.1
Dependency-maintenance patch release.
Changes
- Update Ruff from 0.16.6 to 0.17.0 (#1).
- Update actions/checkout from v4 to v7 (#2).
- Update actions/setup-python from v5 to v7 (#3).
- Bump the installable package to 1.0.1. No runtime API or solver changes relative to v1.0.0.
Verification
Audit passed for release commit f7fdbaa:
- 403 tests passed on both Python 3.11 and 3.12; 91.67% src statement coverage.
- Ruff formatting/lint and mypy passed.
- Offline benchmark quickstarts and wheel/sdist build/install checks passed.
- The 1.0.1 wheel was also built, installed and smoke-tested locally on macOS.
This remains experimental research software; the existing scientific and data-license limitations still apply.
Full changelog: v1.0.0...v1.0.1
v1.0.0
Release of the winding-field implementation and reproducible artifact benchmark.
- Continuous winding-field solver, physical-coordinate constraints and coherence tools.
- Hash-bound candidate evaluation, separation/collapse controls and optional matched CT overlays.
- Offline synthetic benchmark demo with clearly labeled synthetic pictures.
- README CT illustration with attribution, Audit/coverage/license badges, and fixes for NumPy 2.5 chart labels and explicit AFV configuration.
- Installation metadata, documentation, CI and a 90% statement-coverage floor.
Validation: 403 tests passed with 91.67% src coverage on Linux with Python 3.11 and 3.12. Ruff, mypy, offline benchmark demos and wheel/sdist build/install checks passed in GitHub Audit for commit 4fc673e. Local macOS validation also passed.
This is experimental research software. Numerical convergence and visual agreement do not certify physical sheet identity. Real datasets, weights and research-run artifacts are external inputs/outputs; the documented README illustration retains the underlying data terms.
The original v0.1.0 tag, release and assets are preserved unchanged.
v0.1.0
First public release. Wrap-adjacency graph plus a five-method winding-assignment negative result for PHerc 0332 (Scroll 3), with a full checksummed evidence pack (docs/evidence/) and the 2026-07-10 sign-convention correction.
Release assets (SHA-256 in docs/CHECKSUMS.txt):
- wfield_multi.npz — the exact Scroll 3 field input the method scripts consume
- wfield_final_signfix_m1..m4.npz — the four sign-corrected output fields behind the corrected numbers
All five regenerate from public data with the scripts in this repo; the assets just save you the solver runs.