benchmark still crashes on code-only --no-cluster graph in v0.9.27 (follow-up to #1198) #2212

Summary

graphify benchmark <graph.json> still crashes with KeyError: 'links' on Graphify 0.9.27 when the graph was generated by the documented code-only, no-cluster extraction path.

This appears to be the unresolved mode-specific case behind #1198. That issue was closed on the premise that graph.json always uses links, but graphify extract --code-only --no-cluster currently emits edges.

Environment

• Graphify: 0.9.27
• Python: 3.11.15
• NetworkX: 3.6.1
• OS: Ubuntu 26.04 on WSL2 (Linux 6.18.33.2-microsoft-standard-WSL2, x86_64)
• Target corpus: public pallets/itsdangerous, commit 672971d66a2ef9f85151e53283113f33d642dabd

Minimal reproduction

root="$(mktemp -d)"
git clone https://github.com/pallets/itsdangerous.git "$root/target"

HOME="$root/home" XDG_CACHE_HOME="$root/home/.cache" \
  uvx --from graphifyy==0.9.27 graphify extract "$root/target" \
  --code-only --no-cluster --out "$root/out" --force

python - <<'PY' "$root/out/graphify-out/graph.json"
import json, sys
print(sorted(json.load(open(sys.argv[1]))))
PY

HOME="$root/home" XDG_CACHE_HOME="$root/home/.cache" \
  uvx --from graphifyy==0.9.27 graphify benchmark \
  "$root/out/graphify-out/graph.json"

Observed result

The extractor succeeds:

wrote .../graph.json — 218 nodes, 577 edges (no clustering)

The graph's top-level keys are:

['edges', 'hyperedges', 'input_tokens', 'nodes', 'output_tokens']

benchmark exits 1:

File "graphify/benchmark.py", line 107, in run_benchmark
    G = json_graph.node_link_graph(data, edges="links")
...
KeyError: 'links'

Expected result

benchmark should consume every graph.json format emitted by Graphify itself, including the code-only, no-cluster path.

Likely cause

At v0.9.27, graphify/benchmark.py still does:

G = json_graph.node_link_graph(data, edges="links")

The mode used above emits edges. Graphify's affected.py already contains a compatibility normalisation for this exact format split:

if "links" not in raw and "edges" in raw:
    raw = dict(raw, links=raw["edges"])

Applying the same loader normalisation in benchmark.py, preferably through one shared graph-loading helper, should make the command accept both Graphify-generated representations without changing the writer contract.

Prior issue

• benchmark crashes: KeyError 'links' on networkx 3.x (graph.json uses 'edges' key) #1198 was closed after concluding that Graphify always writes links.
• The current reproduction demonstrates that this is not true for extract --code-only --no-cluster --out ... in v0.9.27.

Local evidence

• Captured traceback: /tmp/graphify-v0927-verify/benchmark.log
• Traceback SHA-256: 99cf6f5c86f901e0c465d422c40765beb078eeba488b41ad07f06e21d94248cb