From bedf32e07ca0e556cb2257c0193dd60d31fbda38 Mon Sep 17 00:00:00 2001
From: rajashidattapy <rayan05rio@gmail.com>
Date: Tue, 11 Aug 2026 20:09:59 +0100
Subject: [PATCH] fix(wiki): count each incident edge once in the audit trail
 (#2635)

---
 graphify/wiki.py   | 23 +++++++++++++++++------
 tests/test_wiki.py | 35 +++++++++++++++++++++++++++++++++--
 2 files changed, 50 insertions(+), 8 deletions(-)

diff --git a/graphify/wiki.py b/graphify/wiki.py
index cb9c6cf3..d283552a 100644
--- a/graphify/wiki.py
+++ b/graphify/wiki.py
@@ -73,12 +73,23 @@ def _community_article(
     top_nodes = sorted(nodes, key=lambda n: G.degree(n), reverse=True)[:25]
     cross = _cross_community_links(G, nodes, cid, labels, node_community or {})
 
-    # Edge confidence breakdown
-    conf_counts: Counter = Counter()
-    for nid in nodes:
-        for neighbor in G.neighbors(nid):
-            ed = edge_data(G, nid, neighbor)
-            conf_counts[ed.get("confidence", "EXTRACTED")] += 1
+    # Edge confidence breakdown, over every edge INCIDENT to the community (the
+    # cross-community ones included — those are disproportionately the uncertain
+    # edges, and AMBIGUOUS is what ARCHITECTURE.md flags for human review).
+    #
+    # ``G.edges(nbunch)`` reports each incident edge exactly once. Walking
+    # ``nodes x G.neighbors`` instead visited an edge once per endpoint inside
+    # the community, so an intra-community edge was counted TWICE while a
+    # crossing one was counted once. Intra-community edges are overwhelmingly
+    # the high-confidence EXTRACTED ones — that is what makes a community — so
+    # the split was biased towards confidence and understated the review
+    # burden. On a MultiGraph this also counts parallel edges individually
+    # rather than collapsing them to the first (``edge_data``), which is the
+    # same understatement: an AMBIGUOUS edge parallel to an EXTRACTED one used
+    # to be invisible here.
+    conf_counts: Counter = Counter(
+        d.get("confidence", "EXTRACTED") for *_, d in G.edges(nodes, data=True)
+    )
     total_edges = sum(conf_counts.values()) or 1
 
     sources = sorted({G.nodes[n].get("source_file") or "" for n in nodes} - {""})
diff --git a/tests/test_wiki.py b/tests/test_wiki.py
index 289a0a2b..18013da6 100644
--- a/tests/test_wiki.py
+++ b/tests/test_wiki.py
@@ -96,11 +96,42 @@ def test_community_article_shows_cohesion(tmp_path):
 
 
 def test_community_article_has_audit_trail(tmp_path):
+    """Each incident edge is counted exactly ONCE (#2633).
+
+    The Parsing Layer (n1, n2) touches two edges: n1-n2 (EXTRACTED, both
+    endpoints inside) and n1-n3 (INFERRED, crossing out). Counting
+    ``nodes x G.neighbors`` visited the intra-community edge from both ends and
+    the crossing one from only one, reporting EXTRACTED 2 (67%) / INFERRED 1
+    (33%) — biased towards confidence, which understates the AMBIGUOUS review
+    burden the section exists to surface.
+    """
     G = _make_graph()
     to_wiki(G, COMMUNITIES, tmp_path, community_labels=LABELS)
     parsing = (tmp_path / "Parsing_Layer.md").read_text()
-    assert "EXTRACTED" in parsing
-    assert "INFERRED" in parsing
+    audit = parsing[parsing.index("## Audit Trail"):]
+    assert "- EXTRACTED: 1 (50%)" in audit, audit
+    assert "- INFERRED: 1 (50%)" in audit, audit
+    assert "- AMBIGUOUS: 0 (0%)" in audit, audit
+
+
+def test_audit_trail_counts_parallel_edges_individually(tmp_path):
+    """On a MultiGraph each parallel edge is its own row in the split (#2633).
+
+    Collapsing them to the first (``edge_data``) hid an AMBIGUOUS edge behind an
+    EXTRACTED one running between the same pair — the same understatement as the
+    double-count above.
+    """
+    G = nx.MultiGraph()
+    G.add_node("n1", label="parse", file_type="code", source_file="parser.py")
+    G.add_node("n2", label="validate", file_type="code", source_file="parser.py")
+    G.add_edge("n1", "n2", relation="calls", confidence="EXTRACTED", weight=1.0)
+    G.add_edge("n1", "n2", relation="references", confidence="AMBIGUOUS", weight=1.0)
+
+    to_wiki(G, {0: ["n1", "n2"]}, tmp_path, community_labels={0: "Parsing Layer"})
+    audit = (tmp_path / "Parsing_Layer.md").read_text()
+    audit = audit[audit.index("## Audit Trail"):]
+    assert "- EXTRACTED: 1 (50%)" in audit, audit
+    assert "- AMBIGUOUS: 1 (50%)" in audit, audit
 
 
 def test_god_node_article_has_connections(tmp_path):