Loading...
Searching...
No Matches
MeshKit 0.2 completion validation — 2026-10-29/30

Recorded in the OpenSWMM GUI repository, where MeshKit lived as libraries/meshkit until it moved to this repository (2026-11-13). Commands and evidence paths below are relative to that repository.

All five follow-on items are implemented: native mixed planning and source/size preparation, representative shared-contact checks, a complete six-million-scale mixed job, Python bindings/packages, and the authorized Apache-2.0 license. Measurements below are local Apple arm64/macOS observations, not portable speed or worst-case terrain guarantees.

Gate Result
Qt-free Release build and tests 6/6 CTest suites pass; Python suite contains 8 cases
Optional Qt facades and parity 13/13 CTest suites pass
Integrated GUI meshing worker 97 passed, 0 failed, 3 optional workloads skipped
Quality/settings/SMS GUI checks 7 passed each, including resizing and transition values above 10
Installed native CMake consumer Qt discovery disabled; generates and independently audits a mesh
Clean wheel Python 3.12/macOS arm64 installs; mixed meshing and MKB round-trip run without NumPy, Qt or GDAL
Installed wheel with NumPy All 8 Python tests pass, including buffer lifetime, strided input, metadata, cancellation and terrain
Source archive Fresh extraction builds a wheel with no build-time network fetch
Native dependencies Python extension links only libc++ and libSystem; predicates are compiled in with their separate notice
Local GUI preview Ad-hoc signature verified; separate application launched and native main-window state confirmed

Complete mixed capacity job

The installed release wheel generates 7,123,030 vertices and 7,983,065 cells: 6,250,000 quads and 1,733,065 triangles. Job time is 18.788 s, wall time 18.789 s, peak process RSS 1,862,631,424 bytes (1.86 GB decimal). This includes 100 required mapped patches, triangle corridors, one hole, authored line junctions, reversed/shared line owners, a native bilinear terrain surface and final independent constraint/size/edge/transition/terrain checks. All violation and unknown-terrain counts are zero; 414 source segments pass the delivered audit. No interface retry or partial result is needed. Default patch limits remain active; the probe explicitly permits 12 million vertices, 14 million cells and 8 GiB estimated workspace. Actual peak RSS includes Python, NumPy and retained job/result buffers.

Reproduce with the installed Python package and NumPy:

python libraries/meshkit/tests/mixed_scale_probe.py --full --report /path/to/report.json

The pilot without --full uses the same checks on a small domain. This benchmark excludes DEM reprojection, project adoption, GUI rendering and export. Earlier separate weighted-coarsening and streaming capacity measurements remain recorded below. They must not be added together or described as one end-to-end job.

Supplied Bellinge geometry and terrain

The existing Bellinge mesh supplies the actual 286-vertex footprint (34,892,400 square metres approximately), with two reversed line owners sharing its first boundary segments. The supplied SRTM raster is reprojected once into the footprint's metric frame through an external GDAL data adapter. The native package receives an owned regular grid, with explicit repeated edge nodes and unchanged NoData holes. Its certificate concerns that bilinear reference, not the physical accuracy of SRTM.

The complete mixed job returns 778,423 vertices / 793,141 cells (762,880 quads, 30,261 triangles), using six automatic patches. Time is 1.545 s (1.568 s including Python setup/polling), peak process RSS 558,792,704 bytes. All size, edge, transition, terrain and unknown counters are zero; all 290 delivered source segments pass. Terrain tolerance is 0.5 Z units, maximum nominal size 50 XY units, minimum target 2, adjacent ratio limit 4. The actual largest adjacent ratio is 2.8423. Data preparation and output writing are excluded from that job time. This footprint has no holes; hole handling is exercised in the complete scale job and existing GUI fixtures.

Source SHA-256:

  • DEM: 0ad7561201833ac10bc11b08c58920dbde69b94070c6ce52cad4124451e2fec7
  • Mesh: 5245cce5684acf74cc71246b3b87db0ba3ef50ede485976ed38d31632aae14c9

The first automatic candidate exceeded the patch budget and correctly fell back to triangles. Automatic candidates are now tiled with clearance and refinement headroom before planning, producing the verified mixed result above. A source cleanup regression also caught legal endpoint-on-edge contacts being treated as shortcut crossings; endpoint contacts now remain protected while allowing redundant interior vertices to be removed. Existing crossing, exterior-cell and filled-hole negative cases continue to fail as intended.

Release scope and evidence

License/NOTICE/third-party notice inclusion is checked in the wheel and source archive. Relicensing follows the explicit rights-holder authorization in RELICENSING.md. Python output arrays retain native ownership and are read-only. Progress is bounded, callbacks execute on the caller thread, and cancellation requests do not cross into Python inner loops. Final reports include resolved options, provider identity, bounded decisions and examples. The macOS numeric I/O fallback avoids requiring the newer floating from_chars runtime and preserves subnormal/17-digit values in round trips.

CI configuration adds native build/test/package jobs for Linux, Windows and macOS. Only macOS execution is verified here; the other runners have not been run locally. Native source paths must already express intended intersections as nodes; exact shared/T/overlapping contacts are accepted without proximity welding. The GUI retains its crossing preparation, GIS/project capture and rendering adapters. General arbitrary all-quad decomposition is outside the supported mapped/fallback contract.

Raw evidence is in tests/output/meshkit-completion-2026-10-29/: native-tests-final.log, qt-tests-final.log, gui-pipeline.log, gui-quality.log, gui-settings.log, gui-sms.log, mixed-scale-release.json, realdata-report.json, wheel-clean-smoke.log, wheel-full-tests.log, sdist-clean-build.log, consumer-run.log, cli-smoke.log, and preview.json. Package artifacts are in its dist/ subfolder. Earlier failed trial logs remain as diagnostic history, not acceptance evidence.

Historical validation checkpoints

The following entries describe successive earlier increments and their then-current scope.

Initial extraction validation — 2026-10-29

  • Standalone Release configure/build/install succeeds with Qt 6.9.3 on Apple arm64/macOS 27.0.1, without GUI/engine/GDAL dependencies.
  • Five standalone suites pass: runtime events, constraint simplification/audits, CDT, generator, and generation/native/SMS codec roundtrip.
  • The runtime-only build passes with Qt disabled.
  • A separate installed-package client links meshkit::runtime, meshkit::core_qt, and meshkit::io_qt, generates a mesh and serializes its sections.
  • GUI application build succeeds. Eleven geometry/codec/log suites plus infiltration integration pass. The full terrain suite has 75 passing cases and three normally skipped opt-in workloads; the synthetic scale workload was then run separately.
  • A 600-hole/24-breakline fixture retains 19,249 cells and the baseline hash 665bcd25f66fd6bd.
  • Kernel grid probes complete at 6,000,000 cells and, separately, 6,000,000 vertices / 11,990,002 cells. Their measured peak RSS is 413.5 MB and 766.1 MB respectively.
  • The synthetic-plane terrain worker completes 6,499,882 triangular cells in 16.924 s at 2.450 GB peak RSS, with no terrain unresolved/unknown cells or cap. This excludes rendering, old GUI meshes and final file publication; it is not a worst-case terrain or hydraulic benchmark.

The source-controls/shared-chain increment additionally passes the full integrated pipeline suite (80 cases, zero failures, three optional workloads skipped), native source/preview widget checks (five cases), and native full-dialog checks (six cases). It fixes the coarsening-checkbox defect where triangle-only regions were incorrectly protected as mapped quad interiors. Both terrain-off and terrain-on checkbox-to-worker regressions remove two free vertices, changing the fixture from 287 to 283 cells while retaining constraints.

New capacity observations, including independent audits:

  • Six million vertices on one straight source: 2.567 s simplification, 4.422 s including audit, 719.2 MB peak RSS.
  • Two reversed owners sharing six million source vertices total: 4.188 s simplification, 7.819 s including audit, 809.4 MB peak RSS.
  • A six-million-cell constrained grid: 0.970 s delivered constraint audit after a 0.286 s conversion to the current GUI mesh DTO; 16.860 s total including triangulation, 1.047 GB peak RSS with both representations resident. The audit checks every cell's incidence and four required sides. It does not represent a complex terrain or GUI-render workload. Enable this probe with MESHKIT_AUDIT_DELIVERED=1 meshkit_scale_probe 2001 1501.

Raw evidence and native layout captures are under the parent repository's tests/output/constraint-review-2026-10-29/.

Verified sizing increment

Six standalone suites pass, including independent nominal-size/face-ratio checks, minimum-edge requirements, invalid connectivity/callbacks, cancellation/resource limits, legacy migration, narrow-region footprint bounds, float-floor consistency, and bounded local transition repair. Native settings checks pass five cases, including SI/foot unit persistence and dock resizing; the full native dialog suite passes six cases. The full integrated pipeline passes 82 cases with three optional workloads skipped. Existing legacy behavior and coarsening regressions remain covered.

The curved-terrain/fine-region fixture completes with 391 cells, no terrain unresolved/unknown cells, no global/footprint size exceedances and no transition failures (largest adjacent nominal ratio 1.49977 against 1.5). Repair inserts 25 vertices in two passes. A fixed oversized quad remains intact: verified mode reports its conflicts and strict mode refuses the candidate. A short protected source conflicting with a strict edge floor is refused before meshing.

The sizing audit covers 6,000,000 cells / 8,996,500 shared faces in 2.153 s. With CDT, current mesh DTO and both delivered audits, total time is 7.731 s at 1.423 GB peak RSS. At 6,000,000 vertices / 11,990,002 cells / 17,980,005 shared faces, sizing takes 4.843 s; total 17.855 s at 2.639 GB peak RSS. These uniform grids exercise capacity, not a realistic terrain/GIS/render workload. Six million conservative footprint queries against 1,000 narrow region bboxes take 1.379 s using the region BVH and field minimum pyramid. Source bboxes can overrefine nearby ground.

Evidence: tests/output/sizing-review-2026-10-29/. The GUI uses two repair passes and at most 200,000 additional repair vertices. Fixed quads/mixed interfaces may retain reported conflicts; strict sizing refuses unresolved requested limits. Sizing audits do not certify arbitrary cell overlap or replace the terrain audit.

The full requested implementation record, commands, raw evidence links and remaining work are in plans/workpans/2026-10-29-meshing-implementation-progress.md in the parent repository. That folder follows the repository's existing ignore rules. These observations are local measurements, not portable time/memory guarantees. The final Qt-free geometry API, streaming codecs, licensing review and numerical policy redesign remain outstanding; see README and PROVENANCE.

Unified planning increment

Seven standalone suites pass. Planner regressions cover canonical full shared sides with junction retention, winding/projected origins, explicit priorities, holes/lines/pins/masks, terrain repair/unknown coverage, cancellation/budgets, partial-side refusal, exact domain stitching, required interface refinement, preferred fallback with its source outline, and fixed-patch pin refusal. The full GUI worker suite passes 90 cases with three optional workloads skipped. The new open-block case uses one triangulation pass and has zero delivered upper-size, transition or terrain failures. Native settings and dialog suites pass six and seven cases respectively.

The opt-in planning-only probe completes 6,000,000 mapped cells in 100 patches, 180 shared sides, in 2.372 s at 774,455,296 bytes peak RSS. It excludes CDT, DEM, publication and GUI rendering. Default individual patches are capped at 250,000 cells; full shared logical sides are supported, unmatched partial sides require source decomposition. Mapped planning is limited to four logical sides. Required fixed patches cannot be restationed. These limits are explicit failures/fallbacks, not silent geometry relaxation. Evidence: tests/output/cell-planning-2026-10-29/.

Resolution-weighted thinning/coarsening increment

All eight standalone suites pass, including scalar/version/queue checks, metre/foot normal/removal invariance, narrow full-footprint size protection, pins, unknown/error-bound refusal, directional terrain targets without nominal isotropic refinement, and fixed/preferred weighted quad gates. The complete GUI worker suite passes 91 cases with three optional workloads skipped. Native settings and full-dialog suites each pass seven cases, including widened controls, navigation sizing, transition rates above 10, policy persistence and inactive terrain deactivation. The original coarsening-checkbox cases now verify that delivered vertex count decreases by the reported removals, with terrain off and on; inactive generated CDT slots are omitted from delivery.

The curved-reference/fine-region fixture changes from 286 cells at constant tolerance to 493 with weighting, while removing seven redundant vertices. Applied allowance is 0.01–0.04 m; maximum conservative bound/allowance is 0.999987 with no terrain unresolved/unknown cells and a passed delivered constraint audit.

The indexed queue handles six million insert/update/pop entries in 2.406 s, 120,000,000 bytes queue/index allocation and 133,955,584 bytes process peak RSS. This is a queue-only capacity measurement.

The complete weighted synthetic-plane worker delivers 6,433,024 cells / 3,220,202 vertices from a 24,039,409-pixel source in 43.574 s, peak RSS 2,970,304,512 bytes. One attempt per free vertex removes 38,822 vertices in 12.846 s; peak queue 3,251,646, priority queries 3,376,768. Every delivered vertex is checked for cell incidence. The independent sizing audit covers 9,645,847 shared faces and satisfies all requested size/edge/ratio limits; largest nominal ratio 1.77121 against 3. Terrain allowance is about 0.0025 m with largest conservative bound/allowance 0.999857, no unresolved or unknown cells and no refinement cap. This includes terrain preparation, meshing/coarsening and final audits, excluding GUI rendering/final publication. It is a synthetic-plane capacity observation, not complex real-terrain, hole/line-density, rendering, physical-accuracy or portable timing certification. The complete worker case has 3.22 million vertices; older six-million-vertex CDT probes remain separately scoped.

Evidence: tests/output/thinning-review-2026-10-29/ in the parent repository.

Native jobs, streaming codecs and shared vertices — 2026-10-29

The native-only configuration passes runtime, native job/kernel and native I/O suites. The full standalone configuration passes all 11 suites. A separate installed-package client consumes only core, io, runtime with Qt discovery disabled, including when the installed package also contains optional Qt targets. The CLI generates/inspects a native rectangular mesh. Exact native/Qt kernel parity covers ordinary refinement, resolution-aware coarsening and projected XY.

Native I/O checks include sparse IDs, mixed cells, labels/markers/aliases/units, forward references, multi-record node strings, clockwise normalization, unknown cards, legacy material tags, corruption/truncation/trailing data, record/count limits, cancellation and failed-write preservation of the original destination. The GUI SMS file adapter streams and surfaces compatibility warnings.

Shared constraint regressions exercise exact corners, T-junctions, reversed and partly overlapping lines, domain seams, union membership and identical generated stations in opposite directions. The delivery audit still rejects exterior cells and filled holes. The real GUI worker runs six shared cases across preserve and bounded policies. Full terrain/pipeline suite: 97 passed / 0 failed / 3 optional probes skipped; Quality, settings dock and SMS reader: seven passes each.

Synthetic native triangle job: 6,006,578 cells / 3,006,756 vertices, 7.19411 s, 1,244,659,712 bytes peak RSS, no maximum-size conflicts. Input is a uniform lattice of mandatory source vertices, with no DEM/coarsening/mixed planning/rendering.

Native I/O fixture: 6,002,500 vertices / 5,997,601 quadrilaterals. MKB: 390,020,533 bytes, write 1.24745 s, read/validate 1.43851 s, peak RSS 577,994,752 bytes. SMS: 489,509,490 bytes, write 2.21841 s, read/validate 9.26698 s, process peak RSS 822,935,552 bytes (original mesh retained during read). Temporary capacity files are deleted. These are local synthetic observations, not a realistic end-to-end terrain/GIS or portable performance certificate.

Evidence: tests/output/native-api-review-2026-10-29/ in the parent checkout. The native job's automatic mixed capability remains false; quad planning and source/size-field construction still use the transitional Qt API. See NATIVE_API.md. Python bindings and relicensing are not delivered.