Fix panic on degenerate vertical/horizontal segments in BBox::crossing

[taddyb]

Summary

BBox::crossing() was panicking with unreachable!("vertical line doesn't exit box") when polygon traversal landed in a cell where prev_original and the current coordinate formed a vertical (or horizontal) segment lying entirely outside the cell on the perpendicular axis. The panic killed the entire process rather than failing the affected polygon.

Triggered in production by a MERIT polygon (one of 346K) against the AORC ~0.0083° grid:

thread '<unnamed>' panicked at extractrs-core/src/bbox.rs:165:17:
internal error: entered unreachable code: vertical line doesn't exit box

This is a latent bug inherited from the exactextract port — see isciences/exactextract#126 where the same condition surfaces with low-quality polygon data.

Root cause

Cell::take() calls BBox::crossing(prev_original, c) when c is classified as Outside the cell. prev_original is coords[pos-1] (the previous uninterpolated polygon vertex) used for floating-point robustness — but it can be far from the current cell when the traversal walks across many cells. If prev_original.x == c.x (exactly vertical, e.g. duplicate vertex or true vertical edge) AND that x is strictly outside the cell's [xmin, xmax], the segment doesn't intersect the cell at all, and the vertical-line branch's exit checks both fail.

Fix

Two layers:

1. Upstream guard in Cell::take() — skip prev_original if it would form a degenerate collinear segment outside the cell, fall back to last_coordinate (which is always on or inside the cell boundary, so crossing() will compute correctly).

2. Defense in depth in BBox::crossing() — replace the two unreachable!() panics with a graceful fallback that returns the box side closest to the segment. Any future caller violating the precondition gets wrong coverage for one polygon instead of crashing the process.

Tests

Added 4 new unit tests in bbox.rs exercising the degenerate case in all four directions (vertical left/right, horizontal above/below). Pre-fix these would have panicked. All 64 extractrs-core tests pass; clippy clean.

What this does NOT fix

The traversal logic in intersection.rs::traverse_cells() may still place segments in cells they don't actually pass through (root cause of how we end up calling crossing() in the degenerate state in the first place). Coverage results for degenerate polygons may be slightly off. A follow-up issue should audit that traversal logic — but for now this fix prevents process crashes, which was the immediate blocker.

Test plan

• cargo test --workspace passes
• cargo clippy --package extractrs-core -- -D warnings clean
• Rebuild Python wheel and re-run the AORC extraction that triggered the panic
• Confirm the affected polygon now produces a result (even if approximate) instead of crashing

🤖 Generated with Claude Code