Summary

ford_fulkerson allocates edge_to with network.node_count() but indexes it with NodeIndexable::to_index(...). On StableGraph, deleted nodes leave holes: node_count() counts live nodes, while to_index() uses sparse node indices.

Current behavior

A valid StableGraph with a removed node can panic inside the augmenting-path search when the highest live node index is larger than node_count() - 1.

Concrete regression test

This is small enough to add directly as a crate regression test in crates/petgraph/tests/ford_fulkerson.rs:

use petgraph::{
    algo::ford_fulkerson,
    prelude::StableDiGraph,
};

#[test]
fn test_ford_fulkerson_stable_graph_node_hole_regression() {
    let mut graph = StableDiGraph::<(), u32>::new();
    let source = graph.add_node(());
    let removed = graph.add_node(());
    let sink = graph.add_node(());

    graph.add_edge(source, sink, 1);
    graph.remove_node(removed);

    let (max_flow, _) = ford_fulkerson(&graph, source, sink);
    assert_eq!(1, max_flow);
}

This leaves a valid directed StableGraph with live node slots 0 and 2, so node_count() == 2 while to_index(sink) == 2.

The test fails with index out of bounds: the len is 2 but the index is 2.
After the fix, the test should pass and report a max flow of 1.

Root cause

The algorithm sizes node scratch state in the compact live-node domain but accesses it in the sparse index domain.

Suggested fix

Allocate edge_to from network.node_bound() rather than network.node_count().