Skip to content

Releases

v1.6.4

Choose a tag to compare

@LimeGS LimeGS released this 21 Nov 10:28
958342b

Adds run --field-scope supported (with --field-prior): solves and stores only field chunks with observations and holds the rest at the prior (#83). Core pinned to LimeGS/winding-field#18 (solve_blocks partial_order). Docker security baseline refreshed (#84).

v1.6.3

Choose a tag to compare

@LimeGS LimeGS released this 21 Nov 00:14
ad0395f

Fixes: parallel --level-workers no longer race on the CT chunk ledger (#78); --field-prior accepts a prior whose frame differs only in CT reader metadata, e.g. a coarse field from an earlier release (#79).

v1.6.2

Choose a tag to compare

@LimeGS LimeGS released this 20 Nov 17:22
0922f86

Adds run --field-prior: seed and gauge a fine --roi tile from a completed coarse field (#76).

v1.6.1 — Legacy checkpoints and parallel-level fix

Choose a tag to compare

@LimeGS LimeGS released this 20 Nov 07:35
efc4198
  • Read field checkpoints with float64 deltas written by earlier releases (#73); fixes 'oversized checkpoint payload' on older runs such as PHerc 175A.
  • Fix a journal race when computing native levels in parallel (--level-workers) (#72).

Full changelog: v1.6.0...v1.6.1

v1.6.0 — Flatboi SLIM everywhere, lscm-slim default, faster native surfaces

Choose a tag to compare

@LimeGS LimeGS released this 20 Nov 00:25
b39ecbd

Flattening

  • SLIM is Flatboi in every mode (#55); batches bitwise reproducible (#58).
  • New --flattening-solver lscm-slim (global LSCM seed -> Flatboi SLIM), default for export-sheets on Linux x86_64; lscm and slim remain, each for --scope patches|sheet (#65).
  • All flattening code in winding_field.flattening, old module names kept as aliases (#56).
  • export-sheets SLIM/patches no longer require --app-run (#62).

Performance

  • Native surfaces ~37% faster per level: boundary overlap fast path, vectorized Flatboi OBJ read (#66).
  • Opt-in --level-workers N computes W levels in parallel, same outputs (#67).
  • Passive OpenMP waiting so concurrent runs share a host (#60).
  • Field caches and preparation pool (#43, #46).

Reliability and tooling

  • Restored --nonconvergence-policy and --reuse-from (#47).
  • Numerical core installed from git at its pinned commit (#61).
  • Verification loop: tools/verify.py, REVIEW.md, CI with the base verifier (#59, #64).
  • CI and container test fixes (#52, #53, #54, #57).

Core pinned to winding-field 1.5.4 source (numerically identical to core 1.6.0).

Full changelog: v1.5.8...v1.6.0

v1.5.8 — Configuration, field advice and resilient cancellation

Choose a tag to compare

@LimeGS LimeGS released this 18 Nov 00:48

winding-field-cli v1.5.8

Changes since v1.5.7

  • Complete typed public CLI/configuration entry points, including prepared workflows and export tools; support safe YAML defaults with explicit flag precedence.
  • Add opt-in globally verified field convergence with integrity-checked generations and recovery.
  • Add opt-in consultative field advice, coherent deterministic samples and checkpoint-bound bounded diagnostic history. Advice does not add sweeps, trigger verification, stop calculations or certify physical accuracy.
  • Align the immutable core archive, provenance and CPU/CUDA locks with stable winding-field 1.5.3.
  • Restore the full immutable AVF dependency revision and guard metadata/provenance/lock coherence.
  • Handle the macOS native-worker exit/signalling race during STOP cancellation, escalation and exceptional cleanup without hiding permission errors for live workers.
  • Use runner DNS when building the container test image.

Verification and limits

  • The preceding main revision passed both remote test and container workflows:
    https://github.com/LimeGS/winding-field-cli/actions/runs/44200529027
    https://github.com/LimeGS/winding-field-cli/actions/runs/44200529049
  • The rebuilt CLI 1.5.8 wheel with the actual hash-verified stable core 1.5.3 archive passed 167 focused installed-package tests outside the checkout: distribution metadata/source identity, advisory history/integration, convergence, native runtime and real STOP/SIGINT cancellation/resume.
  • Core release preparation passed 283 focused tests. Built both wheel and sdist; verified packaged metadata/provenance, module bytes, unchanged scientific source hashes and unchanged non-core dependency/lock entries.
  • No local full-suite rerun, Docker rebuild, GPU/real-volume qualification, PyPI/container-registry publication or deployment is claimed.
  • Existing installed runtimes, runs and checkpoints are untouched. Runtime identity checks remain enabled; cross-version checkpoint migration/resume is not promised.

v1.5.7

Choose a tag to compare

@LimeGS LimeGS released this 17 Nov 11:25

v1.5.7

Fix

  • Use one block-local physical geometry for fibre selection, clipping and constraint stencils. Fixes a one-ULP boundary mismatch reproduced with PHerc0483A and other origin/spacing combinations, without changing core 1.5.2 or weakening support checks.
  • Recovery guidance: preserve old runs/checkpoints; after separate authorization, restart only field from verified private copies of preparation/fibres. Do not rewrite implementation hashes or adopt an unqualified partial checkpoint.

Local import performance

  • Reuse bounded decoded/support caches and batch durable ledgers/exports.
  • Exact vectorized clipping, direct calibrated Lasagna reads and bounded ordered import workers (--import-workers, default 1; 0 admitted auto).
  • Synthetic matched whole-pipeline medians improved 37.5% (base-grid) and 45.4% (calibrated), with identical logical outputs. Production scaling remains unqualified; small standalone exclusion-heavy AFV imports incur documented extra cold-start overhead.

Qualification and boundaries

  • This release's installed macOS/Python 3.12 wheel passed 208 focused tests outside the checkout against published pinned core 1.5.2, including recorded source hashes.
  • The runtime changes previously passed 2,466 source tests / 61 skips for throughput plus the focused source/installed boundary regressions and real diagnostic replay: 1,193 curves, 26,943 pairs, none discarded.
  • CI was not awaited by user instruction. No full-suite rerun, Docker rebuild, Linux/container or Python 3.11 execution is claimed for this release.
  • No Vast runtime, original input, output or checkpoint was changed. No recovery or deployment was performed.
  • Full CLI/config parity is a separate subsequent implementation, not part of this release.

v1.5.6 — Correct pinned-core physical boundaries

Choose a tag to compare

@LimeGS LimeGS released this 16 Nov 23:34

Fix

Pin winding-field core1.5.2 to fix the two stride16/172.8 µm field pipelines that failed against the published core1.5.1. Physically valid final-node coordinates can normalize one rounding step past the grid limit; core1.5.2 validates physical bounds first, preserving strict rejection of genuinely outside points and existing support masks. No tolerance expansion or copied solver.

Update provenance and both CPU/CUDA Docker archive hashes. Add installed-core version/source-hash checks to prevent a modified editable dependency from giving misleading qualification results.

Verification

97 critical core tests pass. Installed published-core archive plus CLI1.5.6 wheel, outside checkout with asserted module paths: 76 passed, 1 skipped, including all four real imported-stride pipelines and the two previously failing cases. Independent critical reviews found no actionable Critical/Important issues. Published core archive and all recorded source hashes verified.

Full suite and CI wait omitted at user request; local Docker daemon unavailable, so no Linux/container rebuild pass is claimed. This is a GitHub source release, not a host deployment or PyPI/Docker Hub publication. Existing v1.5.5 is unchanged.

Core release: https://github.com/LimeGS/winding-field/releases/tag/v1.5.2
CLI PR: #33

v1.5.5 — Imported AFV/Lasagna strides 16 and 32

Choose a tag to compare

@LimeGS LimeGS released this 16 Nov 22:42

Changes

  • Support --stride 16 and --stride 32 when both --fibres and --lasagna-input are provided. Other routes retain strides 1..8; inference calibration and bounded-preparation restrictions are unchanged.
  • Plan and prefetch only sampled CT chunks for the combined imported route, instead of the dense native ROI envelope.
  • Preserve selected-level/ROI calibration, physical density, support masks/crop gaps, exact world-coordinate fibre clipping and integrity-bound resume.

Verification

35 new tests passed, including four actual prepare/fibres/field/mesh pipelines at 172.8 µm field spacing, selected CT level, shifted ROI, anisotropic spacing, interrupted/completed resume preservation, mutation rejection, sparse HTTP downloads and sample budgets. 120 distinct critical regression cases passed (CLI module rerun passed after a transient local ENOSPC). Independent critical review found no actionable Critical/Important issues.

Full suite and CI wait were explicitly waived by the requester to prioritize immediate publication. This is a GitHub source release, not a Docker/PyPI publication or server deployment.

PR: #32

v1.5.4

Choose a tag to compare

@LimeGS LimeGS released this 16 Nov 17:41

Changes

  • Stabilize native partial-raster and checkpoint-resume tests with controlled timing instead of wall-clock sleeps.
  • Preserve real geometry, CT rendering and retained-state assertions; production time limits remain unchanged.
  • Align CLI package and runtime versions at 1.5.4.

No numerical algorithms or dependency pins changed. Existing v1.5.3 remains unchanged.

This is a GitHub source release, not a new Docker Hub image or PyPI package. It does not establish GPU qualification or scientific flattening-benchmark acceptance.

Verification

  • Local full suite after the test fix: 2364 passed, 61 skipped, one known fork warning.
  • Focused checks after the version bump: 103 passed, 3 skipped.
  • Review found no issues.

Published at the maintainer's explicit request without waiting for GitHub CI; final CI results are not confirmed at publication time.

Source: 8861e3b4f2d1c0631eb0fcc6deb947a44952faa4, merged through PR #31.