Guard map path sampling against invalid dlon/dlat to prevent macOS pscoast crashes

[Copilot]

The macOS CI failure came from pscoast_JE2 crashing in gmtlib_lonpath after an invalid path-sampling step produced a pathological allocation size. This change hardens map path generation when projection edge cases yield non-finite or non-positive dlon/dlat.

• Root cause

  • gmtlib_lonpath / gmtlib_latpath computed allocation sizes from GMT->current.map.dlat / GMT->current.map.dlon without guarding invalid values.
  • In azimuthal singularity scenarios, this could cascade into impossible allocation requests and segfault.

• Code changes

  • Added defensive checks in:
    • src/gmt_map.c::gmtlib_lonpath
    • src/gmt_map.c::gmtlib_latpath
  • When dlat/dlon is non-finite or <= 0, both functions now return a safe 2-point fallback path instead of entering adaptive sampling/allocation logic.

• Behavioral impact

  • Normal projections/path sampling are unchanged.
  • Degenerate projection-step cases no longer attempt invalid huge allocations; they degrade safely to a minimal path.

if (!isfinite (GMT->current.map.dlat) || GMT->current.map.dlat <= 0.0) {
    gmt_M_malloc2 (GMT, tlon, tlat, 2U, NULL, double);
    tlon[0] = tlon[1] = lon;
    tlat[0] = lat1;
    tlat[1] = lat2;
    *x = tlon;
    *y = tlat;
    return (2ULL);
}

[joa-quim]

We don't have yet a AI policy but I think PRs submitted by agents are not acceptable. Every PR should be submitted by an human.