The check for a constraint that shrinks a curve to a point took under 1e-6 for a point. One solve could squeeze a 49 mm side to about 1.2e-6 mm and the next take it the rest of the way, so neither solve tripped it (two rectangles' corners made coincident, found by a local sketch fuzzer). Under 1e-5 now counts: in 4000 random sketches the constraints that still collapsed a line went from 9 to 0, and every refusal rate stayed the same. Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>