Validating that bend_radius is strictly larger than half the mode plane size

[momchil-flex]

The case when the two are equal could actually be particularly problematic as the bend center lies on a grid point, which is a singularity in our coordinate transformation.

So maybe I should even leave a bit of buffer extra?

[weiliangjin2021]

as for extra buffer, you mean
np.abs(mode_spec.bend_radius) <= plane.size[radial_ax] / 2 * (1-FP_EPS)

[momchil-flex]

Yeah something like that, I'm also wondering if it should even be larger than FP_EPS though. Basically in the trasnform here, the jacobian has a division by zero if the radius lies on a grid coordinate (either primal or dual). So to avoid some very large values, might be better to require it to be even a bit further away.

[weiliangjin2021]

The safest is probably to use those np.nextafter (e.g. in this function

tidy3d/tidy3d/components/geometry/utils_2d.py

Lines 18 to 21 in 3dc27f1

	def increment_float(val: float, sign) -> float:
	"""Applies a small positive or negative shift as though `val` is a 32bit float
	using numpy.nextafter, but additionally handles some corner cases.
	"""

from @dmarek-flex )