Replace recursive FillRecursive with flat FillGrid that tiles along primary
axis, then perpendicular. Extract FindPlacedEdge, BuildRemainingStrip,
BuildRotationSet, FindBestFill helpers. Use array-based DirectionalDistance
to eliminate allocations in FindCopyDistance and FindPatternCopyDistance.
Simplify FindSinglePartPatternCopyDistance to delegate to FindCopyDistance.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Vector implements IEquatable<Vector> with proper GetHashCode for HashSet usage.
Polygon.FindCrossing uses bounding-box pruning to skip non-overlapping edge pairs.
Helper.DirectionalDistance deduplicates vertices via HashSet, sorts edges for
early-exit pruning, and adds a new array-based overload that avoids allocations.
PartBoundary sorts directional edges and exposes GetEdges for zero-alloc access.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Linear phase shows "Linear: 12/36 angles, 45° = 48 parts" with a
running count. Pairs phase shows "Pairs: 8/50 candidates, best = 252
parts" tracking the best result seen so far. Reports on every
completion so the UI always reflects current state.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Parallel loops were flooding the UI with per-angle/per-candidate reports
faster than WinForms could render them. Use Interlocked timestamp checks
to report at most every 150ms, keeping descriptions readable.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Don't overwrite the Detail label with phase-level reports — let the
per-angle and per-candidate descriptions from the parallel loops remain
visible. Only clear the label on completion.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
ReportProgress was not setting Description, so the Detail row always
showed the default em-dash. Now each phase report includes a meaningful
description, and UpdateProgress always updates the label (resetting to
em-dash when null).
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
- Mark _loadAttempted as volatile for correct double-checked locking
- Add Content item to copy Models/ directory to output for ONNX inference
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Adds PerimeterToAreaRatio (perimeter / area) as a spacing sensitivity
indicator for ML feature extraction.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Add FeatureExtractor for computing geometric part features (convexity,
aspect ratio, circularity, bitmask) and BruteForceRunner for generating
training data by running the fill engine and recording results.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Calculate unused plate area by subtracting total part area from the
work area instead of relying on FillScore.UsableRemnantArea, which
could over-report available space.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Replace shape-list iteration with ShapeProfile.Perimeter in both
Part.Intersects and PartBoundary, simplifying the logic and ensuring
only the outermost contour is used for collision detection.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
- Add verbose per-file and per-sheet-size console output during collection
- Skip already-processed parts at the sheet-size level instead of all-or-nothing
- Precompute best-fits once per part and reuse across all sheet sizes
- Clear best-fit cache after each part to prevent memory growth
- Save best-fits in separate bestfits/ zip entries instead of embedding in nest.json
- Filter to Keep=true results only and scope to plate sizes in the nest
- Set nest name to match filename (includes sheet size and part count)
- Add TrainingDatabase with per-run skip logic and SQLite schema
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Run pairs phase first to establish a baseline before linear and
rect-best-fit phases. Replace IsBetterFill with direct FillScore
comparison. Simplify FillPattern to sequential iteration, reusing a
single FillLinear engine instance.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Compute min/max bounds in a single pass alongside part area
accumulation, avoiding the separate GetBoundingBox() call and
redundant iteration.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Store hull edge angles in BestFitResult at evaluation time so they
don't need to be recomputed during the fill phase. Extract
GetHullEdgeAngles(Polygon) overload from FindHullEdgeAngles.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Thread ISlideComputer through BestFitCache → BestFitFinder →
RotationSlideStrategy. RotationSlideStrategy now collects all offsets
across 4 push directions and dispatches them in a single batch (GPU or
CPU fallback). Also improves rotation angle extraction: uses raw geometry
(line endpoints + arc cardinal extremes) instead of tessellation to avoid
flooding the hull with near-duplicate edge angles, and adds a 5-degree
deduplication threshold.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
ISlideComputer abstracts batched directional-distance computation so GPU
implementations can process all slide offsets in a single kernel launch.
GpuSlideComputer uses ILGPU with prepared edge data (precomputed inverse
deltas and min/max bounds) and caches stationary/moving buffers across
calls. GpuEvaluatorFactory exposes a singleton factory method.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Brings in the full NFP implementation: ConvexDecomposition, NoFitPolygon,
InnerFitPolygon, NfpCache, BottomLeftFill, SimulatedAnnealing optimizer,
and INestOptimizer interface. Resolves conflicts by keeping master's
progress reporting infrastructure alongside the new AutoNest methods,
and adapting RunAutoNest_Click to use NFP AutoNest with async/cancellation.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
"Length" is more natural than "height" for flat plate materials.
Renames the field on OpenNest.Geometry.Size, Box.Height property,
and all references across 38 files.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
- Change while(true) to bounded for-loop (max 100 plates) in MainForm
- Use Drawing.Name comparison instead of reference equality for quantity deduction
- Add Math.Max(0, ...) guard to prevent negative quantities
- Tune SA parameters for faster convergence (cooling=0.995, minTemp=0.1, maxNoImprove=500)
- Add --autonest flag to OpenNest.Console for CLI-based NFP autonesting
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Implement geometry-aware nesting using No-Fit Polygons and simulated
annealing optimization. Parts interlock based on true shape rather than
bounding boxes, producing tighter layouts for mixed-part scenarios.
New types in Core/Geometry:
- ConvexDecomposition: ear-clipping triangulation for concave polygons
- NoFitPolygon: Minkowski sum via convex decomposition + Clipper2 union
- InnerFitPolygon: feasible region computation for plate placement
New types in Engine:
- NfpCache: caches NFPs keyed by (drawingId, rotation) pairs
- BottomLeftFill: places parts using feasible regions from IFP - NFP union
- INestOptimizer: abstraction for future GA/parallel upgrades
- SimulatedAnnealing: optimizes part ordering and rotation
Integration:
- NestEngine.AutoNest(): new public entry point for mixed-part nesting
- MainForm.RunAutoNest_Click: uses AutoNest instead of Pack
- NestingTools.autonest_plate: new MCP tool for Claude Code integration
- Drawing.Id: auto-incrementing identifier for NFP cache keys
- Clipper2 NuGet added to OpenNest.Core for polygon boolean operations
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
The push algorithm's copy distance formula (bboxDim - slideDistance)
produced distances smaller than the part width when inflated boundary
arc vertices interacted spuriously, causing ~0.05 unit overlaps between
all adjacent grid parts.
Two fixes applied:
- Clamp ComputeCopyDistance to bboxDim + PartSpacing minimum
- Use circumscribed polygons (R/cos(halfStep)) for PartBoundary arc
discretization so chord segments never cut inside the true arc,
eliminating the ChordTolerance offset workaround
Also parallelized three sequential fill loops using Parallel.ForEach:
- FindBestFill angle sweep (up to 38 angles x 2 directions)
- FillPattern angle sweep for group/pair fills
- FillRemainingStrip rotation loop
Added diagnostic logging to HasOverlaps, FindCopyDistance, and
FillRecursive for debugging fill issues.
Test result: 45 parts @ 79.6% -> 47 parts @ 83.1%, zero overlaps.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
When the greedy maximum-row tiling leaves a thin remainder, tries removing
the last row and re-filling the larger strip. Picks whichever total is
higher. Fixes cases where e.g. 4 rows + 11 remainder = 47 beats 5 rows = 45.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Pre-size TilePattern result list from estimated copy count to avoid
resizing. Remove redundant Program.BoundingBox() call and UpdateBounds()
in MakeSeedPattern — the Part's BoundingBox is already correct.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
max(aUpper_i - bLower_j) across all pairs simplifies to
max(aUpper) - min(bLower), computed in a single pass.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
FindCopyDistance and FindPatternCopyDistance were cloning entire Parts
(including deep Program copies) just to get offset locations for
GetLines. Compute offset locations directly instead. Also skip the
Pattern wrapper in TilePattern — clone parts directly.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
