Merge remote-tracking branch 'origin/master' into claude/avalonia-port-next-tasks-f1f6b6

docs/geometry/directional-slides.md

@@ -14,6 +14,7 @@ Open/incomplete chains and ambiguous contacts conservatively block. Loops are re
 - 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.
+- The best-fit pair search (`RotationSlideStrategy`) starts each slide from an offset on its step grid, then slides every resting position sideways, both ways, and keeps the new resting positions as extra candidates. A pair that locks in a corner (two interlocking channels) therefore closes to the part spacing on both axes instead of stopping up to a step short. This relies on leaving an existing contact being legal.
 - 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 each axis order starts with a coarse box pass. Each candidate then alternates geometry pushes until movement is negligible (at most 20 iterations). The coarse pass remains useful for avoiding sawtooth/rung traps when the starting boxes do not overlap. A box push also stops at an obstacle whose box is beside the path but nearer than the part spacing, so the coarse pass cannot park the copy inside a neighbor's spacing. A push ignores a part the copy starts on, so a candidate can end clear of that part yet inside its spacing: a candidate that keeps the part spacing (measured on `NativeContours`, as the spacing review measures it) always beats one that does not, and a named order that ends too close falls back to the other order. When no order keeps the spacing the copy still lands too close; this is not overlap repair.