Oh, also, the "set" code should be nested in geometry_set.modify_geometry_set([&](GeometrySet &geometry_set) {...
That way the nodes can act on instances as well, and even instances nested in other instances, etc.

• The Set Radius node shows up twice in the search.

I think this is because what I suggested, adding the node to both categories. So looks like that won't work here.
Maybe we could use separate "Set Point Radius" and "Set Curve Radius" nodes? That would help the defaults too actually, for point clouds, 0.05m is usually the default, but for curves something like 1m probably makes more sense.

Also, the set handle position node could use a bit of special behavior with handle types that we couldn't do easily before. It could change them for the selected points like we do in edit mode:

• Vector -> Free
• Auto -> Align

I think this is because what I suggested, adding the node to both categories. So looks like that won't work here.

Think we can also just fix the search so that it doesn't add the same node twice with the same props.

Ok, I'm fine with having two nodes for the radius.

Personally I don't think "Curve" is necessary in "Set Curve Handle Positions", since it's already in the curve category.

The Curve Handle Positions node should be a bit wider by default, so that the name fits in the header.

Do we have a rule for when the term "Curve" and when "Spline" is used?

The Set Spline Cyclic node has a different name in the search vs node header.

In D12687#332952, @Jacques Lucke (JacquesLucke) wrote:

Do we have a rule for when the term "Curve" and when "Spline" is used?

I've tried to make it when it specifically pertains to the spline domain to use spline, otherwise use curve.

I'd like to get an okay for this tomorrow, but I think it might be clearer if the input nodes are also in their corresponding categories, but with separators like this:

def curve_node_items(context):
    if context is None:
        return
    space = context.space_data
    if not space:
        return
    if not space.edit_tree:
        return

    if geometry_nodes_legacy_poll(context):
        yield NodeItem("GeometryNodeLegacyCurveSubdivide")
        yield NodeItem("GeometryNodeLegacyCurveReverse")
        yield NodeItem("GeometryNodeLegacyCurveSplineType")
        yield NodeItem("GeometryNodeLegacyCurveSetHandles")
        yield NodeItem("GeometryNodeLegacyCurveSelectHandles")
        yield NodeItem("GeometryNodeLegacyMeshToCurve")
        yield NodeItem("GeometryNodeLegacyCurveToPoints")
        yield NodeItem("GeometryNodeLegacyCurveEndpoints")

    yield NodeItem("GeometryNodeCurveToMesh")
    yield NodeItem("GeometryNodeCurveResample")
    yield NodeItem("GeometryNodeCurveFill")
    yield NodeItem("GeometryNodeCurveTrim")
    yield NodeItem("GeometryNodeCurveLength")
    yield NodeItem("GeometryNodeCurveSplineType")
    yield NodeItem("GeometryNodeCurveSubdivide")
    yield NodeItem("GeometryNodeCurveSample")
    yield NodeItem("GeometryNodeCurveFillet")
    yield NodeItem("GeometryNodeCurveReverse")
    yield NodeItemCustom(draw=lambda self, layout, context: layout.separator())
    yield NodeItem("GeometryNodeCurveParameter")
    yield NodeItem("GeometryNodeCurveHandleTypeSelection")
    yield NodeItem("GeometryNodeInputTangent")
    yield NodeItem("GeometryNodeInputCurveTilt")
    yield NodeItem("GeometryNodeInputCurveHandlePositions")
    yield NodeItem("GeometryNodeInputSplineResolution")
    yield NodeItem("GeometryNodeInputSplineCyclic")
    yield NodeItem("GeometryNodeSplineLength")
    yield NodeItemCustom(draw=lambda self, layout, context: layout.separator())
    yield NodeItem("GeometryNodeSetCurveRadius")
    yield NodeItem("GeometryNodeCurveSetHandles")
    yield NodeItem("GeometryNodeSetCurveTilt")
    yield NodeItem("GeometryNodeSetCurveHandlePositions")
    yield NodeItem("GeometryNodeSetSplineResolution")
    yield NodeItem("GeometryNodeSetSplineCyclic")

I like the separators, it makes things clearer. IMO the input category also needs this to separate nodes like Position Normal Index from nodes like Object Info Collection Info etc. But this perhaps is more relevant to T91542

I just fixed merge conflicts in this patch, so I'll commit it. Thanks for the work Johnny. I think the code is pretty straightforward, and it's consistent with what we discussed before, so I'm removing Jacques as a blocking reviewer.