[mdsumner]

Fixes #25, the C++ side of hypertidy/controlledburn-rs#6. Same change as hypertidy/controlledburn-rs#18.

Before: a cell was overcounted where a polygon edge passes exactly through a cell corner, or where two edges nearly coincide. #25's ring covered 6.6016 for an area of 6.1578 (cell (12, 18) burned 1.0, exact 0.0943). The other rings from controlledburn-rs#6 burned 0.739 for 0.455 and 0.839 or 1.0 (depending on the start vertex) for 0.339. A sliver of area 4e-13 covered 11 whole cells at 2x resolution. In approx mode, an empty cell burned when two edges both passed 1e-13 right of its centre.

Causes:

• An edge through a corner gave a crossing an ulp off the corner, and the next cell a sliver traversal along its wall where the ring only touches the corner.
• exactextract's left_hand_area depends on the order of the traversal ends around a cell. Rounding can swap two nearly coincident ends (that sliver, or a ring doubling back along itself), and it then returns the complement.
• A single traversal lying along a cell wall doesn't split the cell, but the single-traversal formula treats it as if it did. A ring tracing the grid's bottom edge out and back put one direction in the bottom row and the other in the padding row, and the bottom-row cells came out fully covered.
• A zero-width spike out to a corner vertex and back left a coverage of 1e-16, and that record blanked a fully covered cell.
• Approx compared the rounded intercept with the centre.

After:

• The walker exits through the corner when the crossing is within the lattice tolerance of it (1e-9 of a cell).
• Cells crossed by several traversals, and a single traversal along the walls, get their coverage from Green's theorem, the identity behind controlledburn-rs's CoverageWindow. Each traversal adds its integral of (x - xmax) dy, plus the cell width times the covered length of the left wall, read from the ring's traversals to its left in the row. No ordering is involved.
• A coverage below the sweep's 1e-6 tolerance gets a record only when the ring passes through the cell.
• Approx decides the side of a centre with an exact orientation test: FMA-split products summed as a Shewchuk expansion, with a float filter first.

Golden tables (controlledburn-golden.json lives in controlledburn-rs; dumped from this build): the 28 existing cases match bit for bit, except bowtie coverage. Its closing edge lies on a column wall, and it gains the six cells it was missing: 26 cells before, the exact 32 now. lines_multi's line records come out in a different order from the golden file, as they already do on main. Three cases are added: corner_pass, sliver_coincident_edges and approx_centre_tie. corner_pass is also a new shared fixture (14 now); the CSVs are byte-identical between the repos.

How checked:

• ctest 2/2. test_burn adds five tests, which fail 124 checks on main.
• R: 384 expectations passed, 0 failed (test-lattice-snap.R adds four tests).
• Python: pytest 133 passed (test_regressions.py adds four).
• The Rust port of the same change passes its property suite: 20 of 20 runs at 10,000 cases and 20 of 20 at 100,000. Unfixed main failed 20 of 20 at 100,000.

Status of the other ports (wxylib m20): the hole fix (controlledburn-rs#11, here #27/#31) and the wall-cell fix (#29 and 1a597eb, via #32) are both already on main, and Rust and C++ agree on their golden tables. CoverageWindow/RowPrefix don't exist in the C++ core yet; this PR doesn't port them.

Also noticed: NEWS.md doesn't mention the hole or wall-cell fixes, though 0.3.2 contains both. Not changed here. DESCRIPTION goes to 0.3.2.9000 for the development NEWS entry.

🤖 Generated with Claude Code