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.
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.
aj