validator: judge road-link gaps by lane-boundary contact, not reference-line distance

[kosuke55]

Problem

geom.road-link-gap judged two linked roads by a proxy: the distance between their reference-line endpoints, minus |laneOffsetA - laneOffsetB| as slack.

The proxy breaks whenever the lane offset varies with s, or the two roads meet at an angle. On a measured real map, all ten of its reports were false positives — the lane boundaries in those cases touched to 0.000 m. It is the same failure mode as the soderleden 66 m case, just with the sign the other way.

Fix

Measure what the rule claims to measure. At each contact end, build the road's lane-boundary cross section in world coordinates:

• centre line = reference line shifted laterally by <laneOffset> evaluated at that station,
• plus the running sum of lane widths outward along the reference-line normal (-sin hdg, cos hdg) — left lanes in +t, right lanes in -t.

The gap is the minimum distance between road A's cross section and road B's. Lane offset is handled by construction, so the slack term is gone.

Details:

• the lane section covering the contact station is the one used (multi-section roads),
• <width> evaluation picks the record covering the station, so sOffset spans are respected,
• a link with no contactPoint tries both of the target's ends and takes the nearer, rather than inventing a defect the document never asserted.

Comparing minimum distance asks whether the two sections touch, not whether they are congruent. That is deliberate: lane counts and widths differ across a link at merges, ramps and junction connectors (91 of 156 links on one measured map), so requiring congruence would trade one class of false positive for another. A road whose lanes are wholly displaced from its neighbour's is still reported.

Default threshold is 0.3 m, down from 0.5, now that the measurement is direct rather than a proxy with slack.

Evidence

• False-positive pin: a synthetic fixture where road B's reference line is pushed 5 m aside by a laneOffset ramping along s while the lanes meet exactly — the old rule warned, the new one reports nothing.
• Detection pin: the same construction with the lanes genuinely 1.5 m apart still warns, and the message reports the lane distance.
• Further pins for multi-section roads, multi-record <width> spans, differing lane counts, a missing contactPoint, and a caller-supplied threshold.
• Mutation matrix still 10/10.
• False-positive gate: 0 MAP_DEFECT errors across the 6 checked-in fixtures and the 26-map external corpus.
• soderleden road 7 keeps its warning — its lanes really are tens of metres from the road it declares as predecessor, so the finding survives on lane distance and not just on the proxy.

Suite: 438 passed / 6 skipped, tsc --noEmit clean.