Background knowledge not working

Hi! I have come across an issue where even when I forbid certain nodes when using PC, they still appear in my causal graph result:

from causallearn.utils.PCUtils.BackgroundKnowledge import BackgroundKnowledge
from causallearn.search.ConstraintBased.PC import pc

cg_without_background_knowledge = pc(X)  # Run PC and obtain the estimated graph (CausalGraph object)
nodes = cg_without_background_knowledge.G.get_nodes()

bk = BackgroundKnowledge() \
    .add_forbidden_by_node(nodes[0], nodes[1]) \
    .add_forbidden_by_node(nodes[1], nodes[2]) \
    .add_forbidden_by_node(nodes[0], nodes[2]) \
    .add_forbidden_by_node(nodes[0], nodes[3]) \
    .add_forbidden_by_node(nodes[1], nodes[3]) \
    .add_forbidden_by_node(nodes[2], nodes[3])

cg_with_background_knowledge = pc(X, background_knowledge=bk)

assert cg_with_background_knowledge.G.get_edge(nodes[2], nodes[3]) is None
assert cg_with_background_knowledge.G.get_edge(nodes[0], nodes[1]) is None

I get this error, as well as an error for some other node combinations in my background knowledge:

AssertionError Traceback (most recent call last)
Cell In[9], line 18
15 cg_with_background_knowledge = pc(X, background_knowledge=bk)
17 assert cg_with_background_knowledge.G.get_edge(nodes[2], nodes[3]) is None
---> 18 assert cg_with_background_knowledge.G.get_edge(nodes[0], nodes[1]) is None
20 g, edges = fci(X, background_knowledge=bk)
22 sns.heatmap(g.graph, cmap='coolwarm', annot=True, xticklabels=node_names, yticklabels=node_names)

AssertionError:

It seems like if the background knowledge is being used, it's being overriden somewhere. Please help! This also happens with FCI. I looked at the code and I can't figure out why this could be happening.

Reproduce the assertions with the BackgroundKnowledge example, then trace how background_knowledge is passed through causallearn/search/ConstraintBased/PC.py and the FCI entry point. Check where forbidden node pairs are applied during graph construction and orienting; done means the reported forbidden edges are absent for both PC and FCI, with regression coverage for the example.