The point was more that problems that need to use atan() in a context where the line may point near the Y axis don't flip out over "floating point precision", they just use atan2().

Problems like the one in the grandparent that need to intersect nearly-parallel lines (or subspaces) use special techniques better suited to the problem and don't just feed it to a linear solver.