This was a two-part bug: a user error + API error.
* User Error: before calling bpy.ops.constraint.childof_set_inverse()
for a constraint defined on a bone, you firstly need to explicitly
make that bone the active bone. To do that, you do
armature.bones.active = posebone.bone # or something similar
* API Error: active bone setting was a bit too strict. It only allows
setting the active bone if the new bone comes from the same armature,
but was overlooking the fact that RNA pointers may have been created
through the object using the armature instead.
97002c2d90