Plan Cutting spent most of its time on work that never changed the outcome:
many-holed parts paid quadratic shape-pair and ring-containment costs before
any search, every plate was captured twice, and failing plates burned their
whole budget in exhaustive fallback passes that never found a plan on 127
production plates.
- OverlapMaterial: skip native shape pairs whose bounds are clear by more
than the query allowance; gate polygon touch and winding tests on vertex
bounds.
- LeadMaterialSnapshot: gate ring enclosure tests on ring extents when
finding the perimeter and rejecting nested cutouts.
- CuttingPlanBatch: capture each plate once; the kept-order retry derives
from the same capture through CuttingPlanSnapshot.WithPreservedOrder.
- JointCuttingPlanSearch: end the kept-order full pass and the sequence-free
recovery and full passes after FallbackStalls (2) stalls' worth of
expansions when programs are regenerated; plates with only fixed programs
keep the exhaustive search so a refusal stays genuine.
On 127 archive plates (lead-ins cleared, 35" scrap cut-offs) total time fell
from 164 s to 102 s with identical outcomes; the problem-nest corpus gives
the same status and expansion count on every plate.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Fine inscribed flattening still let a native arc cut 6.2e-7 into a 1e-5
clamp unseen: less than one chord's sag, so the layout check, interactive
fill and Check Overlaps all called it clear. Tiny-clamp checks now
circumscribe perimeter arcs and inscribe cutout arcs (lines stay exact), so
the polygons hold every bit of real material. Any real overlap is reported;
the cost is a possible report within the tolerance of an arc, never a false
clear. The analyzer builds the same polygons instead of a second
OverlapMaterial read.
PM 407f5eb7-329f-41e4-947e-64ad52aa4034 (delta review s6-delta-review.md)
The tolerance-free rectangle test was fed approximated material: the layout
check passed it the 1e-4-grid validation outline (a strip of real material
that wholly contains a 1e-5 clamp could vanish), and interactive fill and
the overlap check passed 0.001 inscribed arc chords (a clamp inside a disk
but outside its chords read as clear). All three now flatten the native
contours at a quarter of the clamp's smaller side, lines kept exact, so the
arc-chord sliver can never hold the clamp. Ordinary clamps are unchanged.
PM 407f5eb7-329f-41e4-947e-64ad52aa4034
ajandClaude Fable 5.1