ajandClaude Opus 5.5
5eb7fd920b
Six-decimal incremental PEP programs round arc centers, so an arc's end tangent misses its neighbor's by about gap / radius (1.6e-5 rad on the offset 5/32" fillet in the reported nest). Two parts resting one spacing apart then gave the slide contact a Minkowski cone a hair wider than a half-plane, which the classifier read as the whole plane: every direction blocked, so the top part could not be pushed sideways along the part below. SlideContact now reads a cone up to TangentSlack (1e-4 rad) past a half-plane as a noisy half-plane and widens its boundary and concave-ray comparisons by the excess, so near-tangential slides read as tangential. A direction admitted that way can enter material by at most travel x excess at that contact, and every other contact is still classified on its own; wider cones remain real overlap and block. Raising JointTolerance was not needed for this case and is left at 1e-6. Whole-job benchmarks (synthetic nests plus the PEP job at 3 and 60 parts, --parallel 1) are identical before and after for every engine. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
7.0 KiB
Directional slides and repeated pushes
Behavior contract
For initially non-overlapping closed material boundaries, a directional slide stops at the first contact that blocks forward motion. Leaving an existing contact or sliding along a straight shared edge is legal. Skipping a contact must not skip the rest of that obstacle: a later hook or the opposite wall of a hole still stops the slide. Distances retain the existing Tolerance.Epsilon snapping and double.MaxValue no-hit convention; supplied vector directions are unit vectors.
SpatialQuery.DirectionalDistance overloads and CPU best-fit batches use the same event sources and contact resolver. Events carry both contact points in their initial world frames, rather than only a snapped distance. The classifier uses closed-loop material sectors, native-curve containment for hole depth, and the curvature of the supporting boundary at a tangential contact. A full-circle arc has no physical corner at its seam. Raw ray helpers remain first-touch primitives, not material-aware slide queries.
Open/incomplete chains and ambiguous contacts conservatively block. Loops are recovered with ShapeBuilder's Tolerance.ChainTolerance, so a source contour that closes within it (PEP-rounded drawings can miss by 1e-6) stays a closed loop; a stricter closure test would leave the whole part unresolved and freeze any part resting against it. Rounded arc centers also tilt a joint's tangents (about gap / radius), so two parts touching along a shared edge can give a contact cone a hair wider than a half-plane; the classifier reads a cone up to SlideContact.TangentSlack (1e-4 rad) past a half-plane as a noisy half-plane and treats directions within that excess of its boundary as tangential, so a part can still slide along its neighbor. A wider cone is real overlap and blocks. This is not an overlap-repair operation or a general replacement for layout validation. Caller-provided contact topology must represent the same boundaries and offsets as the query. Prepared geometry must not be mutated; prepare a classifier before sharing it between parallel queries. Edge-array queries still sort their arrays, and recover loop order from private copies before classifying contacts.
Callers
-
PlateView uses
SelectionManager.PushSelected→Compactor.Push. The zero-spacing nudge-and-discard workaround is removed. Cutout contours on stationary obstacles are retained, and the plate entry's existing-overlap filter accounts for holes rather than relying solely onPart.Intersects(which compares outer perimeters). -
With Calculated distance checked in Tools > Set Offset Increment, PlateView's arrow keys use
SelectionManager.StepSelected→Compactor.Step. The step is the selection's linear-fill copy distance (FillLinear.CopyDistance) at the plate's part spacing, so a shape that nests into its own copy steps less than its size. That copy distance is the same in both directions along an axis. The move is not a slide: like a fixed nudge it always moves the full step, ignoring the work-area edge, clamps and other parts, because interlocked rows (staggered triangles, say) cannot slide straight past their neighbors to the next row. It can land off the plate or on a part; the layout check reports that as it does for any nudge. -
Linear fill inherits the shared native-entity query unchanged. Extents fill passes complete boundary loops instead of direction-filtered fragments.
-
CPU best-fit batches prepare contact topology once and use all vertices plus curve/line interior and curve/curve tangency events. The old leading-half vertex filter cannot establish the next blocker after a skipped touch.
-
GPU kernels retain nearest-hit reduction and return unsnapped contact witnesses. The shared CPU classifier accepts a blocking witness or replays the full query after a nonblocking witness, preserving tied/later blockers. Both batch APIs honor active buffer lengths and refresh mutated/reused segment arrays. The GPU distance adapter sends only exact cardinal directions to the axis-only slide interface; arbitrary directions and native curves use the shared CPU path.
-
Shift-click while cloning parts settles the copied group toward the plate quadrant. If its starting bounds overlap any placed part's bounds, it skips the coarse bounding-box pass and uses geometry directly; otherwise it tries coarse horizontal/vertical and vertical/horizontal orders. Each candidate then alternates geometry pushes until movement is negligible (at most 20 iterations), and the group nearest the quadrant's work-area corner wins. The coarse pass remains useful for avoiding sawtooth/rung traps when the starting boxes do not overlap.
Regression coverage
SlideContactTests exercises cardinal line, translated line, reused edge-array, arbitrary-vector, native-entity, and both CPU batch paths. Cases include winding reversal, nonzero origins, rotated hooks, holes, separating circles, positive-distance grazing followed by a blocker, full-circle arc seams, concave/straight junctions, thin rings, and circle/line interior contact.
CompactorTests covers the reported sequence (push left with spacing, then right/up/down), genuine zero-distance blocking, zero/nonzero-spacing later hooks, and inside-hole pushes through both direct and plate entry points. Physical spacing is measured from raw outlines rather than the inflated contours used by the solver.
Verification commands:
dotnet test OpenNest.Tests/OpenNest.Tests.csproj --filter 'FullyQualifiedName~SlideContactTests|FullyQualifiedName~CurveContactDistanceTests|FullyQualifiedName~CompactorTests'
dotnet test OpenNest.Tests/OpenNest.Tests.csproj
dotnet test OpenNest.Engine.Tests/OpenNest.Engine.Tests.csproj
dotnet test OpenNest.IO.Tests/OpenNest.IO.Tests.csproj
The isolated repair tree (excluding other sessions' fill-performance and spacing-expander work) passed 130 targeted cases and the full Debug suites: main 1758 passed / 21 skipped, engine 300 passed, IO 41 passed. The main Release suite passed 1728 / 21 skipped. Skips are not counted as passes. The 142 GpuSlideContactTests also pass in a net8 harness linking the production GPU class and test source, using ILGPU 1.5.1's CPU accelerator (not a mocked distance solver). The Windows desktop/test project cross-build passes in Release. Neither physical GPU execution nor Windows UI interaction was runtime-verified on Linux.
Remaining hardening
This repair does not change saved best-fit cache versioning, add a clearance acceptance gate to every fill entry, repair existing overlapping layouts, or claim the earlier real-DXF/grid-validator discrepancies in pair-spacing checks are resolved. That document's measured candidate counts describe its earlier tree; removing CPU projection filtering and adding interior curve/line contacts does not substitute for rerunning its corpus. General Part.Intersects hole semantics remain unchanged outside Compactor. Profile the new classification path before attempting optimizations; retain the first-blocking-contact regressions.