Fix graticule lines vanishing across most projections

[nighthawk]

Why

Whole families of graticule lines were rendering as nothing. Meridians disappeared in Mercator/Equirectangular/Gall-Peters, parallels disappeared at reference longitude 0, and EqualEarth and NaturalEarth were each dropping 18 lines. These looked like separate bugs in separate projections; they turned out to share one root cause.

wrapShift documents that it relies on consecutive points being close together in projected space — but the interpolator deliberately breaks that invariant. It stops subdividing the moment a span projects straight, which is right and cheap for drawing, yet leaves adjacent points arbitrarily far apart wherever the projection is locally linear over a long span. Wrap detection then can't tell "linear over a long span" from "jumped across a seam", and splits lines that never crossed anything.

The fix restores the invariant instead of loosening the heuristic that depends on it. A new Interpolator.densify pass runs after the existing one and bisects any pair further apart than a quarter of the projection size, stopping at the same floor the main pass uses — so a real discontinuity converges to a tight jump straddling the seam, which is exactly the signal wrap detection wants, while a smooth span closes up in a couple of levels.

The subtlety worth knowing about: bisection has to be path-aware. An edge from 170° to -170° means the 20° hop across the antimeridian, so splitting it in raw longitude lands on lon 0 — the long way round — and erases the crossing. An edge spanning a full 360° is the exception, meaning "all the way round", and keeps its long path rather than collapsing to zero.

No API change. The commit message has the mechanics.

Evidence

Measured over a full graticule (36 meridians + 17 parallels) across eight projections at two reference longitudes, diffed against main:

reference	projection	blank lines before → after
0°	Equirectangular / Mercator / GallPeters	17 / 21 / 17 → 0
0°	EqualEarth / NaturalEarth	18 / 18 → 0
0°	Danseiji I	2 → 0
77.5°	Mercator	4 → 0
both	Cassini, Danseiji IV	0 → 0 (byte-identical)

97 previously-blank lines now draw, and no line got shorter anywhere. Real content is barely affected because its vertices are already dense — the EqualEarth continent bench moves 10695 → 10700 vertices.

GraticuleCoverageTests is added as a standing guard. It measures drawn length rather than point count, because the failures that shipped were lines rendering as zero-length pieces: invisible, but not absent enough for a point-count check to notice.

Test plan

Automated: 60/60 pass. The one golden-file update is testBasicSVGGeneration, where M 100,100 L 300,100 gains a collinear midpoint at 200 — same geometry, one extra vertex from densify.

Manual, in Cassini (load Examples/Data/graticule-2.geojson as a layer):

• Cycle through every projection and confirm meridians and parallels both render. Mercator, Equirectangular and Gall-Peters are the ones that were worst.
• Drag the reference-longitude slider through 0°. Parallels used to vanish entirely at exactly 0; they should stay put now.
• Check Cassini specifically — far-hemisphere meridians should still fold around the top and bottom edges rather than drawing straight down the map.
• Confirm continent rendering is unchanged (this touches the shared projection path, not just graticules).

Known, not addressed here

Two pre-existing defects this does not fix, both verified present on main:

• A meridian lying exactly on a bezier projection's seam (reference ± 180°) is discarded wholesale, because pointInPolygon accepts an exact vertex match but treats a point on an edge as outside. Excluded and documented in GraticuleCoverageTests; fixing it changes polygon clipping semantics on a documented hot path, so it wants its own change.
• Danseiji IV shows straight runs at the notches in its outline. Untouched by this change (its output is byte-identical), so it's a distinct cause — most likely the concave multi-crossing path in boundarySplit.

Independent of #43; both branch off main.

🤖 Generated with Claude Code