Same-pair same-relation collision downgrades an EXTRACTED edge to INFERRED and rewrites its provenance

[fa137]

What happened?

When two edges land on the same node pair with the same relation, the guard added for #2803 cannot fire, so G.add_edge() overwrites and the last write wins. Because semantic edges are always concatenated after AST edges, a semantic INFERRED edge silently overwrites an AST EXTRACTED edge and takes its source_file / source_location with it.

Expected: on a same-pair, same-relation collision, the higher-confidence edge should survive, and confidence should never be downgraded from EXTRACTED to INFERRED.

Actual: the surviving edge reports confidence: INFERRED, the semantic confidence_score, and the doc's source_file. Nothing in the output records that a deterministic AST edge existed.

Steps to reproduce

from graphify.build import build_from_json

extraction = {
    "nodes": [{"id": "a", "label": "A"}, {"id": "b", "label": "B"}],
    "edges": [
        # AST edge first — the skill pipeline builds `ast['edges'] + sem['edges']`,
        # so AST edges always come first for the same pair.
        {"source": "a", "target": "b", "relation": "calls", "confidence": "EXTRACTED",
         "confidence_score": 1.0, "source_file": "a.py", "source_location": "L724"},
        # semantic edge second, SAME pair and SAME relation
        {"source": "a", "target": "b", "relation": "calls", "confidence": "INFERRED",
         "confidence_score": 0.95, "source_file": "some_note.md", "source_location": None},
    ],
}
G = build_from_json(extraction)
print(G.get_edge_data("a", "b"))

Reversing the two edges in the list does not help either, because add_edge is still last-write-wins and the caller controls the order.

Error output or graph output

surviving edge: {'relation': 'calls', 'confidence': 'INFERRED', 'confidence_score': 0.95,
                 'source_file': 'some_note.md', 'source_location': None}

Expected:

surviving edge: {'relation': 'calls', 'confidence': 'EXTRACTED', 'confidence_score': 1.0,
                 'source_file': 'a.py', 'source_location': 'L724'}

Observed on a real 2019-node corpus. api/app.py:724 is broker = MockTDAAccount(...). AST extraction emits it correctly:

$ graphify ...  # AST side
api_app_build_app -- calls -> broker_mock_tda_mocktdaaccount | EXTRACTED | loc L724

But graph.json contains only this edge for that pair:

{"relation": "calls", "confidence": "INFERRED", "confidence_score": 0.95,
 "source_file": "graphify-out/memory/query_20260920_232355_....md", "source_location": null}

So GRAPH_REPORT.md reports a relationship stated in plain code as an inference guessed from one of graphify's own internal memory files, and the EXTRACTED fact is unrecoverable from graph.json.

Scale on that corpus: 1 of 3701 edges for this exact case, but 27 pairs total where an EXTRACTED edge was replaced by a non-EXTRACTED one (the other 26 are covered in the follow-up issue below).

Graphify version

0.9.61. Re-verified against the v0.9.65 sources: graphify/build.py L1415-1423 and _GENERIC_RELATIONS (L64) are unchanged, so the defect is present in the latest release.

Environment

macOS 15 (darwin arm64), Python 3.12, uv tool install graphifyy, /graphify invoked from an OpenCode session.

Additional context

Related and distinct:

• Edge collapse picks the surviving relation alphabetically, so references overwrites calls every time (144/144 pairs) and those calls leave the call graph #2803 (CLOSED) added the _GENERIC_RELATIONS guard, but that guard only fires when the incoming relation is generic and the existing one is not. A same-relation collision is not covered by it, and neither is a collision between two relations that are both in the set.
• build_merge() defaults to directed=False, so graphify update silently downgrades a directed graph to undirected — betweenness becomes meaningless with no warning #2342 concerns update and direction, not relation/confidence selection.

Suggested direction: before G.add_edge, compare the incoming and existing edges and keep the stronger fact rather than the later write, e.g. rank by EXTRACTED > INFERRED > AMBIGUOUS, and on an equal-confidence tie prefer the AST-sourced edge over a semantic one. The _GENERIC_RELATIONS logic can then layer on top of that instead of standing in for it.