The optimizations applied to point-comparison math for 2d queries in core.h result in incorrect results. Basically:

return _y > p._y ? p._y + radius >= _y : _y + radius >= p._y;

... returns incorrect results on linux32, but the equivalent :

volatile double sum = _y > p._y ? p._y + radius : _y + radius;
return _y > p._y ? sum >= _y : sum >= p._y;

... returns the correct results. Volatile is required.

From mathias, adding compiler flags:

-mfpmath=sse and either -msse2 or -march=pentium4

... solves the problem, but limits our target architectures.