Operators sometimes pull a clamp for a manual layout, but only Auto Nest's
last-resort search could take clamps off a sheet. Right-clicking a clamp in
Select mode now offers to remove it from that sheet or put it back, reusing
the existing per-sheet RemovedClamps, so the zones, cut-time charge, display
and .nest persistence all behave as they do for Auto Nest removals.
The change is refused while a nest run or fill may be reading the clamps,
sharing the guard the Clamps dialog already used.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The spec review found two places where the marker and threshold
contract does not hold.
A clamp narrower than the chord tolerance is still judged
tolerance-free against finely re-read material and reported as the
whole clamp rectangle. That existing safety policy is kept, so its
marker is the clamp's centre: within half the clamp's size (under
0.0005) of the material, though not necessarily on it. Document this
on PlateOverlapPair.Marker and in the overlap docs, and lock it in the
tiny-clamp region test.
Also document that 1e-8 grid rounding can move an area within
rounding error of the 1e-5 threshold to either side of it.
Far from the origin a grid step smaller than the coordinate's
precision leaves x (or y) unchanged: at 1e16, 1e16 + 1 == 1e16. The
starting grid loop then never ended and allocated cells without limit,
even with the new maxCells budget (a hang dump reached 4.3 GB). Stop a
row or column once its step no longer advances.
Grids that advance are untouched, so the bit-identical differential
against the former sorted-list implementation still passes. Found by
the independent quality review of the Clipper2 overlap branch.
The overlap crosshair sat on the pair's area centroid. For a split,
holed or concave overlap, that point can be off the red material:
between two patches, in a hole, or inside a curve.
PlateOverlapPair.Marker is a point inside the pair's largest island.
It is the island's pole of inaccessibility (PolyLabel), searched to a
tenth of the island's area/perimeter within a 4096-cell budget. If
that bounded search does not land inside (a long thin sliver), the
marker falls back to the midpoint of the widest interior run of a
horizontal scan line placed between vertex heights.
The overlay crosshair, marker hit testing, hover coordinates
("Marker:") and the cutting-inspection ring now use the marker.
Centroid stays the true area centroid. View > Overlap Check > Display
> Centroids is now labelled Markers.
Before the marker existed, its call sites used the centroid. Two
patches, a covered frame, a half annulus and a thin crescent then put
it off the material; now all four land on it. Disabling the scan-line
fallback fails its test. On the 389-plate corpus, markers add about
0.4 ms per overlapping pair; pair and area results are unchanged.
PolyLabel kept its candidate cells in a sorted List, which costs
O(n^2) through Insert and RemoveAt(0). A 48 x 0.002 sliver searched
to 1e-4 took about 27 s.
Use a PriorityQueue keyed by (MaxDist, insertion order). Equal bounds
still pop newest first, and initial cells in sorted order, so every
existing caller gets bit-identical poles. A frozen copy of the
sorted-list implementation checks this on 1800 random shapes and on
symmetric shapes with tied poles. Swapping the tie order fails 25 of
the differential rows.
A new Find(outer, holes, precision, maxCells) overload stops after
about maxCells distance evaluations and returns the best cell so far.
When the starting grid would exceed a quarter of the budget, it is
coarsened so the whole shape is sampled first. The unbounded overload
behaves as before.
The material-overlap diagnostic clipped every triangle of one part
against every triangle of the other, then carved each fragment by
every hole triangle. Curved overlaps produced thousands of fragments,
and a part sitting inside another part's cutout was slow even when
clear.
PlateOverlapAnalyzer now intersects each candidate pair's hole-aware
material once with Clipper2, in part-local coordinates on a 1e-8 grid.
Each operand is its outline minus the union of its holes (positive
fill rule). Regions become connected islands: a counterclockwise
outline plus clockwise Holes, including an island that both parts'
holes enclose together. PolygonAreaMoments.TrySubtract gives each
island's net moments. A pair is reported when its total shared area
exceeds Tolerance.Epsilon. Out-of-range or nonfinite coordinates make
the pair incomplete rather than clear. The overlay and the cutting
inspection highlight fill outlines with their holes.
On 389 real PEP-converted plates the pair, clamp-finding and issue
sets are identical to the previous kernel. Pair areas are equal or up
to 0.0002 larger: the kernel drops fragments below its area threshold.
With material prepared, the per-plate clipping maximum fell from
2.6 s to 52 ms (total 5.0 s to 0.5 s).
Engine and validation checks still use Collision.
PlateOverlapRegion vertices repeat their closing vertex, and the
desktop overlay drops that last vertex when it builds the fill path.
The tolerance-free tiny-clamp finding returned four open corners, so
the overlay drew it as a triangle. Close the ring like every other
region.
The hand-written kernel was kept Clipper-free so it could be ported to
a GPU. That port is on hold in favour of Clipper2, starting with the
material-overlap diagnostic. Engine and validation callers stay on the
kernel until a Clipper2 near-contact differential and the layout gate
pass. The former port contract stays documented for reference.
A direct CutOff.Regenerate(plate, settings) call with no perimeter
cache argument trusted Part.BoundingBox plus a sub-epsilon roundoff
allowance. A legitimately larger arc refit (ConvertProgram fitting
an imported arc's center to its chord) can protrude past that cached
box by more than the roundoff, so an uncached manual cutoff could
emit a segment crossing the part's true converted material.
Every production call site (Plate.RegenerateCutOffs,
AutomaticCutOffPlanner, the desktop cutoff tools) already builds and
passes a Plate.BuildPerimeterCache dictionary, so this only affects a
direct/uncached caller. Extract the per-part conversion out of
BuildPerimeterCache into Plate.BuildPartPerimeter and use it as a
lazy, on-the-fly fallback in CutOff.ComputeSegments whenever no cache
dictionary is supplied at all, so the uncached path gets the same
conservative envelope as the cached one instead of silently trusting
the bare bounding box. Cached behavior (including a cache's own
explicit null entries from a failed conversion) is unchanged.
OpenNest PM 13f7a685-0564-4cbf-8f46-a7805a0288a0.
Expand now reads the plate on a worker while its Cancel window pumps
messages, so a still-running fill could commit parts into the plate the
search is enumerating (InvalidOperationException: collection modified).
Refuse the command while whole-job nesting, a NestProgressForm or a fill
on any view of the nest is active. The usual plate-action guard is not
reused: ActionSelect reports busy whenever parts are selected, which
Expand requires.
Expand Spacing ran its search on the UI thread, so a complex layout froze
the app for minutes with no way out. Run the search on a worker; when it
lasts past 0.3 s, show a small modal window with elapsed time and Cancel.
Cancel, Escape and the close box request cancellation through the
Expander token and the window stays open until the worker stops, so the
layout is never touched afterward. A cancelled run moves no part and
reports it in the status bar instead of the "select two parts" notice.
A cut-off line passing exactly through a vertex of the clearance outline
got duplicate intersection hits and fell back to the part's bounding box.
Automatic strip lines sit on a 0.05 lattice that often lands on such
vertices, and the strip search preferred them because the box fallback
measured as less cutting, so planned cut-offs looked box-trimmed while a
slightly moved manual line was trimmed correctly.
Slice the polygonal clearance outline exactly with half-open edges:
crossing vertices count once, touching vertices zero or twice, edges on
the line are excluded whole. Zero clearance and non-polygon outlines keep
the existing path and its conservative fallback.
Choosing a part from the editor's drawing list armed the pickup only in
Click. Moving the mouse more than 3 px before release started a
drag-and-drop, which takes mouse capture, so the list never raised Click
and nothing attached to the cursor on the plate.
The list now raises DrawingPickedUp on a single left press over a row
and the editor arms the pickup there. Dragging onto the plate still
drops a part at the release point; a completed drop clears the pickup so
a second copy does not follow the cursor. Presses below the last row,
right presses and the second press of a double-click do not pick up.
The Windows run 38614147947 failed ActiveCutOffDragNeedsLeftMotionAndEndsAfterSelectionClears because a screen-space drag to world X=11 rounded to X=10.96. Assert bounded movement, then verify the released coordinate stays unchanged.
Cut-off scrap lines were a solid 1.5px near-black stroke that read like
another sheet/part outline. They now render as a 2px orange-red dash so
operators can spot scrap cuts at a glance, and the manual cut-off preview
is recolored to match the applied line. Selected cut-offs keep their solid
blue highlight.
Add bitmap regression tests (CutOffDisplayTests) that verify the dashed
orange style, the dashed gaps, and the unchanged selection highlight.
PEP-converted nests stored etch marks only in each placed part's locked
cutting program; the drawing program held the contour cuts alone. Because
Remove Lead-ins rebuilds a part from its drawing, clearing lead-ins also
erased the marks from every part.
Recover the marks while reading the nest: when a drawing has no scribe
moves and every placement that kept its saved program carries the same
marks, add them to the drawing in its own frame and refresh that drawing's
clean placements. Saved programs are never rewritten, drawings that already
have marks are untouched, and drawings whose saved programs disagree are
left alone rather than guessed at.
Avoid shortest round-trip decimal text for already-integral doubles. The failing tie regression and the independent 9,163-case exact-integer sweep now keep the fewer-pierce result.
Reserve nominal-line slots in the candidate lattice, compare rounded lengths in exact micro-units before pierce ties, and report real fixed-grid cut length. Per-line notices name local increases instead of negative savings. Failing-first regressions cover dense spacing, decimal ties and the legacy summary.
Check Overlaps dismissed a tiny clamp by the part's coarse material bounds
before the fine tolerance-free test ran, but that material is inscribed at
the 0.001 chord tolerance: real material in the sliver beyond a convex arc
(a short arc cap, a rotated crescent) lay outside those bounds and the clamp
read as clear. For clamps below the tolerance the bounds are widened by the
chord tolerance before rejecting.
PM 407f5eb7-329f-41e4-947e-64ad52aa4034 (review s6-delta3-review.md)
Circumscribing every arc holds all material along convex arcs but drops a
sliver along concave ones (a notch), so a tiny clamp 6e-7 into the material
beside a notch read as clear. Tiny-clamp checks now test the part flattened
both ways (all arcs inscribed, all arcs circumscribed): along any arc one of
the two holds the real material, so together they find every real overlap.
PM 407f5eb7-329f-41e4-947e-64ad52aa4034 (delta review s6-delta2-review.md)
The writer's sentinel rule compared zones only, so an empty sheet following
a rail that differs from the default only in clamp removal minutes was
dropped on save although ChangePlateDefaults treats it as configured stock.
The rule now also requires the sheet to follow the default rail itself.
PM 1469b4e0-dc59-416b-8118-560b726aa1e1
ajandClaude Opus 5.5