tpmslab 0.3.4__tar.gz → 0.3.9.dev12__tar.gz

tpmslab-0.3.9.dev12/CHANGELOG.md

@@ -0,0 +1,376 @@
+# 0.3.9.dev12 — linear tetrahedron export
+
+- Remove the optional 10-node curved export and its COMSOL import validation case.
+  The NAS and NPZ exports contain four-node tetrahedra.
+
+# 0.3.9.dev11 — optional surface envelope; hand-off (2026-11-05)
+
+- `optimize_mesh(..., surface_envelope=mm)` (CLI `--surface-envelope`):
+  interface nodes may leave the exact implicit surface by at most this
+  first-order distance |g|/|grad g| when that improves the local worst quality
+  (fTetWild-style envelope; default None keeps the exact surface). The maximum
+  deviation is checked and reported (`optimization.surface_deviation_max_mm`).
+  With effort="high" and 8.66 µm (fTetWild's default 1e-3 of the box diagonal)
+  over the 16 bcc_crosscompare cases: median solid p01 0.62, median minimum
+  0.48; versus fTetWild on the same boundary p01 higher in 15/16, minimum
+  higher in 15/16, faster in 16/16 (1.4–4.2x), but solid volume error lower
+  in only 4/16 (typically +0.01 to +0.1 pp higher; the envelope moves are not
+  volume-neutral yet).
+- docs/HANDOFF_ZH.md: state, architecture, reproduction commands and the open
+  task list for the next maintainer.
+
+# 0.3.9.dev10 — high-effort optimisation (2026-11-05)
+
+- `optimize_mesh(..., effort="high")` (CLI `--optimize-effort high`, web
+  "高质量优化"): every cell below quality 0.7 is targeted instead of 0.5. The
+  standard threshold capped the BCC solid p01 near 0.52 at every resolution.
+  16 bcc_crosscompare cases (BCC α 0.3 / boundary 0.35): median p01 0.51 →
+  0.57 (up to 0.66), median end-to-end time 2.7 s → 4.1 s. Against fTetWild on
+  the same boundary (single thread): p01 higher in 12 of 16 cases, solid volume
+  error lower in 15 of 16, 2–5× less time; fTetWild keeps a higher minimum in
+  8 of 16 (mostly Diamond and IWP box layers) and higher p01 on the coarsest
+  Diamond/IWP levels.
+- Edge-removal candidates are scored with the same scalar arithmetic but
+  without per-call cache overhead (identical meshes; about 7 % less time).
+
+# 0.3.9.dev9 — box-layer snapping, curved quadratic export, periodic faces (2026-11-05)
+
+- `Config.snap_alpha_boundary` (CLI `--snap-alpha-boundary`): separate snap
+  threshold for lattice vertices on the outer box. Diagnosis on the BCC meshes:
+  80–100 % of the worst 1 % cells touch the box, mostly where the TPMS meets a
+  box face at a shallow angle. With BCC α 0.3 / boundary 0.35 over the 16
+  bcc_crosscompare cases: solid p01 up in 12 and down in 2 (at most −0.013),
+  solid minimum up in 8 and down in 1 (−0.003), volume error unchanged within
+  ±0.05 pp. 0.4 produced a degenerate case (Diamond r10) and is not recommended.
+  Default None keeps dev8 arrays.
+- `make_quadratic(model)` / `save_model(..., quadratic=True)` / CLI
+  `--quadratic`: `mesh_quadratic.nas` with 10-node CTETRA and CTRIA6; mid-edge
+  nodes of implicit-surface edges lie on the exact level set (box planes kept
+  flat, both phases share them). Jacobian sampled at 19 points per element;
+  curvature halved where det J / det J_linear < 0.2. Solid volume error versus
+  the 256³ integral drops 5–64× (e.g. Gyroid r13 0.335 → 0.006 pp); solid +
+  fluid still fill the box exactly. COMSOL import of the 10-node file is not
+  yet verified (evidence/run_dev9_validation.py).
+- `Config.periodic_faces` (CLI `--periodic-faces`): on every axis where the
+  field repeats (checked numerically), opposite faces get identical node sets
+  and triangulations: field samples copied, snaps paired, coordinates enforced
+  and `report["periodic"]` verified after generation and optimisation. The
+  optimiser moves paired face nodes together and skips face retriangulation
+  there (p01 Gyroid 0.511 → 0.505, Diamond 0.504 → 0.492; about 2× slower).
+- `report["wall_resolution"]` for sheets: 2q/|∇f| wall thickness against the
+  lattice spacing, with a warning when the thinnest 5 % is under one cell.
+- Optimiser incidence is built by array grouping: identical meshes, 14 % less
+  optimisation time on the 12-case BCC timing set.
+- Not adopted: thin-wall adaptive edge bisection (benchmarks/refine_experiment.py)
+  did not beat uniform BCC at equal element count (Diamond_low and Gyroid
+  0.08→0.5, P1 k_eff) and was 2–4× slower.
+- Benchmarks: bcc_crosscompare.py, other_tools.py (TetGen, fTetWild, microgen),
+  quadratic_volume.py, bcc_figures.py; validation launcher Run-Dev9Validation.cmd.
+
+# 0.3.9.dev8 — optional body-centred cubic background lattice (2026-11-05)
+
+- New `Config.lattice = "kuhn" | "bcc"` (CLI `--lattice`, web "背景格子").
+  `bcc` adds voxel centres and box-face centres; every pair of face-adjacent
+  voxels contributes four congruent tetrahedra (mean ratio 0.952, dihedral
+  angles 60°/90°), and each box-face square is coned through its face centre
+  (four half tetrahedra, mean ratio 0.800). The background is conforming to
+  the box faces, so clipping, snapping, implicit-root cuts, two-phase
+  generation and `optimize_mesh` are unchanged. Default stays `kuhn`.
+- BCC has about twice the tetrahedra per voxel: use about 5/6 of the Kuhn
+  resolution for a similar element count. Recommended with `snap_alpha=0.3`
+  (web UI uses 0.3 for BCC, 0.25 for Kuhn).
+- `benchmarks/bcc_vs_kuhn.py`: four quality_vs_gmsh designs, Kuhn r12/r16
+  (α 0.25) vs BCC r10/r13 (α 0.3), optimised: solid p01 higher in 6 of 8
+  cases (equal in one, −0.013 in one), minimum higher in 7 of 8, median
+  +0.07 to +0.17 in all, volume error within ±0.07 pp of Kuhn, generate +
+  optimise 1.1–3.8× faster. Larger α (0.4, or 0.25/0.40 by edge length) raises
+  quality further on some designs but increases the volume error.
+- Per-voxel density bookkeeping uses an explicit background-to-voxel owner
+  map instead of `owner // 6`.
+- `benchmarks/itt_*.py`: stage-0 prototype of template inheritance and
+  iso-surface transport (not part of the package; see evidence/itt_stage0).
+
+# 0.3.9.dev7 — less repeated work and phase-aware surface moves (2026-11-05)
+
+- Generation reuses SVD polygon ordering only for shared faces with exactly
+  matching vertex IDs and order. Scalar fan screening rejects clear failures;
+  close decisions and accepted fans retain the original NumPy arithmetic.
+- Cache ordered-tetrahedron metrics during reconnection, up to 65,536 entries;
+  invalidate on every coordinate refresh. The cache preserves signed volume
+  and does not change numerical decisions.
+- Evaluate chord error on actual incident faces, without padded five-dimensional
+  triangle copies; skip candidates already excluded by quality or orientation.
+- Mixed-phase surface moves preserve each phase's local minimum quality.
+  Round-level quality, median, chord and volume budgets remain unchanged.
+  This can retain more rounds and improve final quality, with a design-dependent
+  time cost. Record rejected phase trials in each round's history.
+- Add numerical regressions for scalar screening, cache invalidation, packed
+  chord errors and per-phase trial acceptance.
+
+# 0.3.9.dev6 — generation work reuse and geometry budgets (2026-11-05)
+
+- Generation reuses face incidence for connected components, precomputes face
+  masks, and batches shared-vertex coordinate validation in two-phase meshes.
+- Surface smoothing recomputes hard chord guards against the current geometry
+  of each independent batch, with bounded chunk memory. `optimize_mesh` now
+  defaults to `chord_tolerance=0.0`; explicit `None` retains soft ranking only.
+- Each optimisation round checks per-phase minimum/p01/poor-volume quality,
+  the existing median-loss budget, and cumulative phase-volume drift (default
+  0.3 percentage points of box volume; configurable or `None`). A failed round
+  restores the previous mesh and records its rejection reasons.
+- Recompute density profiles, domain/phase volumes, counts and interface area
+  from the final mesh. Profile `slab_count` records unequal-bin weights.
+- Rebuild current boundaries before optional SLSQP conditioning. Clear input
+  COMSOL solve evidence after optimisation.
+- Add regressions for live-neighbour chord guards, exact rollback, geometry
+  reports, SLSQP boundaries, shared-coordinate validation and face components.
+
+# 0.3.9.dev5 — speed (2026-11-05)
+
+- Generation: identical meshes (mesh_sha256 checked on eight configurations
+  against 0.3.9.dev3), about 30% faster.
+  - `choose_fan` scores its 3-10 cell fans with scalar volume / edge-length
+    arithmetic that reproduces numpy's bit for bit (the 3-term `einsum` order
+    is detected at import against the local numpy; `pow` stays in numpy);
+    falls back to the numpy path if no order matches.
+  - `quality.unique_rows` packs integer rows into one int64 key: identical
+    output to `np.unique(axis=0)`, several times faster; used for face/edge
+    incidence in generation, two-phase partition, conditioning and audits.
+  - `boundary_digest` is vectorised (same stable sorts, same bytes).
+  - Expressions are compiled once (whitelisted AST, empty builtins) instead of
+    walked node by node on every evaluation; values are bitwise identical.
+- `optimize_mesh`: about 2x faster than 0.3.9.dev4 at equal or better quality.
+  - New `tpmslab.batch_smoothing`: interior and surface smoothing move
+    independent vertex sets (stars sharing no cell) together; all trial
+    positions of a batch are scored in one pass with the vertex-affine volume
+    and vertex-quadratic edge-length forms, without materialising the moved
+    tetrahedra.
+  - Adjacency is built once per round and shared; boundary arrays for the
+    surface step come from the mesh bookkeeping; oriented arrays are rebuilt
+    once at the end.
+  - The chord-error criterion uses one |grad g| per candidate vertex (six
+    fewer field evaluations per triangle centroid); cap flips compare
+    area x |g| directly.
+- Mesh results of `optimize=True` change (Jacobi-style batches instead of
+  vertex-by-vertex order); default meshes do not.
+
+# 0.3.9.dev4 — faster, stronger cut-layer optimisation (2026-11-05)
+
+- Local mesh operations score their 1-20 cell stars in pure Python
+  (`quality.tet_scalar`/`star_scalar`) or one vectorised pass over all trial
+  positions (`quality.star_trials`) instead of one numpy call per star;
+  numpy call overhead was ~95% of the dev3 optimiser time.
+- General Shewchuk-style edge removal (`_Mesh.remove_edge`): an interior edge
+  with a ring of 3-7 tetrahedra is replaced by the maximin triangulation of
+  its ring (O(n^3) dynamic programme); 3-2 and 4-4 flips are the n=3 and n=4
+  cases. Candidate edges/faces that failed are not retried within a round.
+- Interior and surface smoothing try two directions per vertex (neighbour
+  centroid and the point that makes the worst incident cell regular) at four
+  step lengths and keep the best admissible one.
+- Each round works on the quality tail (cells below the 5% quantile, never
+  above `quality_threshold`), `focus_quantile=0.05`.
+- Surface moves are chosen with a chord-error criterion: among positions that
+  reach at least `chord_gain_floor` (0.1) of the best quality gain, the one
+  whose incident surface triangles depart least from the level set
+  (area x |g|/|grad g| at the centroids) is taken. An optional hard gate is
+  `chord_tolerance`. Without this, the stronger smoothing moved the phase
+  volume away from the exact implicit volume on coarse meshes.
+- The per-round SLSQP `condition_mesh` pass is no longer part of the default
+  `smooth=True` path (it changed no p01/minimum value in the study);
+  `smooth="slsqp"` keeps it.
+- Tests: edge removal on a five-cell ring, vectorised star trials, and the
+  optimise test now asserts that the solid volume does not move away from a
+  128^3 midpoint integral of the exact field (was: shift < 0.1 pp).
+- Box-face edge removal (`_Mesh.remove_boundary_edge`): an edge lying in a
+  flat box face, whose two box-face triangles are coplanar, is removed with
+  its open ring; the face quad takes the other diagonal. Region, box faces
+  and phase volumes are exactly unchanged.
+- Interface cap flips (`_Mesh.cap_flip`, two-phase, `surface=True` only): a
+  cell with all four vertices on the interface and two interface faces is
+  handed to the other phase (the interface takes the other triangulation of
+  the same four surface points) and its now-interior edge is removed; kept
+  only when quality improves and the new interface triangles are not further
+  from the level set. Interface conformity, triangle counts per boundary id
+  and connectivity are preserved; exterior/interface arrays are rebuilt.
+- Default meshes (optimize=False) are unchanged.
+
+# 0.3.9.dev3 — cut-layer mesh optimisation (2026-11-05)
+
+- New `optimize_mesh` (`tpmslab.optimize`, `generate(..., optimize=True)`,
+  `tpmslab generate --optimize`, browser checkbox): local reconnection of the
+  cut layer (2-3, 3-2 and 4-4 flips restricted to interior faces/edges),
+  maximin Laplacian interior smoothing with SLSQP refinement of the poorest
+  stars, and `tpmslab.surface_smoothing`: boundary vertices slide within the
+  box faces and *along the exact implicit level set* (vectorised Newton
+  re-projection, residual < 1e-9 mm). Exterior/interface connectivity, phase
+  and domain labels are unchanged; phase volumes shift by ~0.01 pp; the
+  result is audited (`quality_audit`, interface, box planes, residuals).
+- `benchmarks/quality_vs_gmsh.py`: like-for-like comparison with Gmsh on the
+  same boundary polyhedron (fixed boundary, interior only) and with Gmsh
+  re-meshing the surface as well; see docs/COMPARISON_ZH.md for numbers.
+- `condition_mesh(max_median_loss=...)` exposes the median-loss budget of the
+  non-regression gate (default unchanged, 5%).
+- `quality.cross` / `quality.signed_volumes`: fast small-array helpers.
+- Default meshes (optimize=False) are unchanged from 0.3.9.dev2.
+- Not yet run in COMSOL: `evidence/run_dev2_validation.py` gained an
+  `optimized_r12` case (flow and elasticity on optimised meshes).
+
+# 0.3.9.dev2 — review fixes (2026-11-05)
+
+Read-only review of 0.3.9.dev1 followed by these changes. No COMSOL run was
+made for this version; see `evidence/README.md` for what the shipped
+evidence does and does not cover.
+
+- Evidence: rebuild `evidence/authorized_smoke_summary.json` from the attempt
+  records (both smoke cases passed on their second attempt; the first-attempt
+  summary is kept as `authorized_smoke_summary.first_attempt.json`); make
+  `evidence/run_authorized_smoke.py` idempotent; add `evidence/README.md`.
+- One COMSOL launcher (`tpmslab.comsol.run_comsol_class`) for the
+  elasticity/flow bridge, `comsol_remesh` and `comsol_audit`: 6.2/6.3 mesh
+  API reflection, isolated `-prefsdir`, short `-tmpdir/-recoverydir/-data`
+  paths, log capture, marker parsing and the COMSOL version string recorded
+  in `run.json`/`comsol_verification.json`. `remesh-quality` previously used a
+  hard-coded `geometricModel("")` that does not compile on 6.2.
+- `volume_components` now counts connectivity through shared faces in both
+  single- and two-phase modes; vertex/edge-only contacts are reported as
+  `pinched_contacts` with a warning instead of counting as one body.
+- Reopen check compares the integrated volume with a relative tolerance
+  (was an absolute 1e-15 m^3).
+- Pre-flight estimate of the background-lattice memory; oversized
+  `cells`×`resolution` fail with an explicit message (limit adjustable via
+  `TPMSLAB_MAX_LATTICE_BYTES`).
+- Two-phase generation samples, calibrates, audits and snaps the lattice
+  once and shares it between the solid and complementary passes.
+- `edge_interpolation="implicit_root"` uses vectorised bracketed bisection
+  started from the linear cut instead of per-edge `brentq`; the
+  unbracketed-cut fallback is unchanged. All other modes produce
+  byte-identical meshes to 0.3.9.dev1 (checked by mesh_sha256 on seven
+  configurations). `implicit_root` meshes differ only on edges where the
+  field grazes zero several times within a few micrometres (12 of 5001
+  nodes, at most 4.3 um, in the n=10 Gyroid check): both versions return
+  a point on the implicit surface, but not always the same one.
+- Two-phase `density_profile` is recomputed from the final (projected) solid
+  cells; `density_profile_method` names the binning.
+- Off-grid sampling probes run for built-in families as well; custom
+  formulas are still rejected on a miss, built-in families get a warning.
+- Flow demo restricts `spf` to fluid domains that connect Z- and Z+ and
+  records `excluded_fluid_domains`; only a model without any spanning
+  channel is rejected.
+- `element_order` (1 or 2) and `processes` (`-np`) options on `build_mph`,
+  `run_comsol`, the CLI and the convergence benchmark; recorded per attempt.
+  Order 2 is not yet covered by any shipped evidence.
+- `generate --mph/--solve` accepts the same launch options as `comsol`.
+- The browser COMSOL path no longer deletes the previous top-level
+  verification file; `build_mph` preserves it inside the new attempt.
+- Documentation: version wording aligned with the 6.2.0.290 evidence.
+
+# 0.3.9.dev1 — isolated COMSOL execution
+
+- Explicit per-attempt preference files, CLI and benchmark options; global preferences unchanged.
+- Short native temporary, recovery and workspace paths prevent Windows path-length failures.
+- Actual COMSOL 6.2 solid and pore-flow smoke checks and documented comparison with related tools.
+
+# 0.3.9.dev0 — Python to COMSOL workflow
+
+- Correct nonuniform three-grid power-law fits and document the revised estimates.
+- Preserve a separate, input-hashed directory for each COMSOL attempt; clear failed current state.
+- Add run_comsol API and a comsol CLI command for saved models and material settings.
+- Fix empty-cut projection and add bounded off-grid probes for custom formulas.
+- Check compiler output as well as exit status; support 6.2/6.3 mesh API signatures.
+- Checkpoint convergence rows, resume studies, fix Gmsh cleanup and label elasticity snapping accurately.
+- Stream hashes; preserve historical delivery records under history/pre_0.3.9.
+- Real COMSOL 6.2 Java compilation succeeded for both smoke cases; NAS import blocked by file-system security preferences. No successful solver validation is claimed.
+
+# 0.3.8.dev0 (local improvement build; not published)
+
+- Add opt-in pre-cut lattice-vertex snapping (`Config.snap_alpha`, `--snap-alpha`): a lattice vertex within `snap_alpha` of a cut point along one of its lattice edges is moved onto the isosurface before clipping, so the near-degenerate sliver that cut would have produced is never created. Default `snap_alpha=0` leaves every existing output byte-identical.
+- Snap onto the exact implicit root by default (`snap_target="implicit"`); `"linear"` keeps the linear-cut point. Outer-box nodes may only move within their box face; moves that would invert a background tetrahedron, flatten a tetrahedron onto one isosurface, or pinch the material or its complement along a lattice edge are reverted and counted in `report["vertex_snapping"]`.
+- Snapping is applied to the full inequality set before any fluid-side sign flip, so complementary solid/fluid passes share identical vertices and cut keys; the interface audit, box-face audit and scalar-diffusion patch check pass on all 12 study designs.
+- Add `benchmarks/snapping_study.py` (12 paired designs, three variants, recomputed statistics, dense-sample reference density and implicit-surface fidelity) and `tests/test_snapping.py`.
+- Allow `snap_alpha` together with `edge_interpolation="implicit_root"` / `"quadratic_root"`: root projection now uses the snapped lattice coordinates and keeps the linear cut when the true field does not bracket a root (`report["edge_root_unbracketed_cuts"]`). With both enabled every interface node lies on the implicit surface; the median solid-density error drops from 1.52 to 0.69 pp (worst 3.66 -> 2.03 pp) while the Q<0.01 count stays at 1 on the 12 designs.
+- Add a browser checkbox (切割前顶点吸附) that sets `snap_alpha=0.25`, restore it from saved parameters, and show snapped/reverted vertex counts in the result notes.
+- Measured on the 12 designs at `snap_alpha=0.25`: aggregate Q<0.01 cells 2710 -> 1 (Q<0.1: 60217 -> 250), tetrahedra 489696 -> 211328, per-phase minimum quality typically 0.001-0.01 -> 0.08-0.18. Costs: median absolute solid-density error unchanged (1.52 -> 1.51 pp) but the worst case rises (3.66 -> 4.03 pp); interface-node RMS distance to the implicit surface rises 10-25 % (10-40 um at 5 mm / resolution 12). No solver-accuracy or speedup claim is made.
+
+# 0.3.7.dev0 (local improvement build; not published)
+
+- Add bounded interior maximin smoothing with analytic quality derivatives and unchanged boundary/interface nodes, connectivity, domains and phase labels.
+- Reconstruct input cell incidence; reject inverted cells, duplicates, nonmanifold faces and missing boundary/interface declarations.
+- Audit candidates independently of optimizer success. Enforce phase minimum/p01/poor-count non-regression, a 5% relative median-loss budget and per-domain volume retention; roll back failed candidates with evidence.
+- Expose `condition_mesh`, `quality_audit`, `generate(..., condition=True)`, CLI flags and a browser checkbox. Optional quality thresholds reject before export.
+- Include 12 paired mesh records with independent geometry and exact scalar-diffusion patch checks. Do not infer FE/CFD accuracy or speedup from geometric quality alone.
+
+# 0.3.6.dev0 (local development source; not a public release)
+
+- Add `combine="blend"` for spatial transitions between two built-in or custom TPMS fields.
+- Expose the secondary family, transition axis and bounds, weight profile, and calibration weight-sample count in Python, JSON, and the local browser interface.
+- Calibrate the blended field using fixed-weight periodic quantiles; preserve exact original fields outside the transition and when both expressions are identical.
+- Use one blended field for complementary solid/pore meshes and root interpolation, and export complete combined-design metadata.
+- Add tests, a Python example, a JSON example, and combined-geometry documentation. The feature adds one composite solid phase and its complementary pore phase.
+- Preserve frozen experiment sources and unchanged default mesh arrays; this feature has not established a mesh-quality or FEM accuracy advantage.
+
+# 0.3.5.dev0 (local experimental source; not a public release)
+
+- Add an optional COMSOL-native `RemeshFaces` plus `FreeTet` workflow for already verified two-phase exports.
+- Require save/reopen, preserved domain and boundary selections, exact cross-phase face incidence, volume and interface-area tolerances, and improved solid quality before accepting the output.
+- Add opt-in `edge_interpolation="implicit_root"` for two-phase meshes. It projects shared cut vertices to implicit-field edge roots subject to local positive-tetrahedron checks; the default remains linear interpolation.
+- Add experimental `edge_interpolation="quadratic_root"`: batch-evaluate the continuous field at three positions on each cut edge, fit a quadratic crossing, and apply the same inversion guard to both phases. It offers a lower-cost approximation to iterative exact roots in the tested graded designs.
+- Keep COMSOL external and optional. Preserve the 0.3.4 direct-mesh benchmark provenance separately from these exploratory modes.
+
+# 0.3.4
+
+- Add configurable density-calibration sampling via `m_cal` (default 56).
+
+- Remove fixed upper limits on per-cell resolution and per-axis cell counts.
+- Use an uncapped integer input in the browser interface.
+- Preserve lower bounds, integer validation and numerical algorithms.
+
+# 0.3.3
+
+- Set the per-cell resolution maximum to 48 in Python and the browser slider.
+- Retain removal of the total background-grid cap.
+- Synchronize documentation and add boundary regression checks.
+- Report the running generator version instead of a stale literal; geometry is unchanged.
+
+# 0.3.2
+
+- Remove the fixed 200,000-background-voxel limit from configuration validation.
+- Retain per-axis cell-count and per-cell resolution validation.
+- Add configuration regression coverage for grids above the former limit.
+
+# 0.3.1
+
+- Use English for Python progress callbacks, exceptions, API messages and exported warnings.
+- Update generator metadata; preserve numerical algorithms and historical validation records.
+
+# Changelog
+
+## 0.3.0
+
+- Package the verified complementary solid/fluid workflow and independent audits.
+- Include density, refinement, interface, channel and performance records with provenance.
+- Retain the validated numerical implementation from commit 686b783e0385.
+- Clarify geometry/solver response stability and local quality-strategy limitations.
+- Add isolated-wheel reference checks, exact-commit release preparation and public version documentation.
+- Preserve historical validation provenance; GitHub publication and DOI registration remain separate records.
+
+
+## 0.3.0rc1 — solid and pore-fluid domains
+
+- Add complementary solid/fluid meshes with shared interface vertices and faces.
+- Preserve independently connected solid and fluid domains, phase arrays and named boundaries.
+- Add fluid preview, domain metadata and optional COMSOL named selections.
+- Verify three fixed-wall pore-flow examples in COMSOL; deformation-coupled FSI is not implemented.
+- Add analytical partition and multi-family interface regression tests.
+
+## 0.2.0rc1 — local release candidate
+
+- Introduce src-layout packaging, public Python API and command-line commands.
+- Keep Flask and the browser interface optional; COMSOL remains optional.
+- Add conservative interior-fan selection with unchanged external boundary.
+- Add quality-tail diagnostics, deterministic mesh/boundary fingerprints,
+  environment records and file-integrity manifests.
+- Normalize configuration lists to immutable tuples and apply phase controls
+  consistently in the field helper.
+- Include reproducible comparisons, automated tests and release documentation.
+
+Earlier rc entries describe local release candidates.

{tpmslab-0.3.4 → tpmslab-0.3.9.dev12}/CITATION.cff

@@ -2,7 +2,7 @@ cff-version: 1.2.0
 message: "Please cite the exact software version used."
 type: software
 title: "TPMS Lab"
-version: "0.3.1"
+version: "0.3.9.dev12"
 authors:
   - family-names: Chen
     given-names: Xin

tpmslab-0.3.9.dev12/MANIFEST.in

@@ -0,0 +1,17 @@
+include README.md LICENSE CHANGELOG.md THIRD_PARTY_NOTICES.md
+recursive-include licenses *.txt
+recursive-include docs *.md *.m
+recursive-include examples *.json *.py
+recursive-include tests *.py
+include CITATION.cff
+include requirements-quality-verified.txt
+include Start-TpmsLab.ps1
+include Start-TpmsLab.cmd
+include Setup-Python.ps1
+include Run-ComsolConvergence.ps1
+include Run-ComsolConvergence.cmd
+include requirements-python-comsol-verified.txt
+include SOURCE_ORIGIN.json
+recursive-include benchmarks *.py
+
+include docs/runtime_hashes_0.3.1.json