graphify cluster-only CLI silently flips directed graph to undirected (v0.4.23)

[johnp3004]

What's happening

Running graphify cluster-only . on a directed graph (directed: true in graph.json) silently re-saves it as undirected (directed: false) and canonicalises edge source/target order to NetworkX's internal storage order, scrambling the original directions.

Reproduction (graphify v0.4.23, macOS, Python 3.14)

# Start with a directed graph.json
python3 -c "
import json
data = json.loads(open('graphify-out/graph.json').read())
data['directed'] = True
open('graphify-out/graph.json','w').write(json.dumps(data))
print('directed:', json.loads(open('graphify-out/graph.json').read())['directed'])
"
# → directed: True

graphify cluster-only .
# → Done — N communities. GRAPH_REPORT.md, graph.json and graph.html updated.

python3 -c "
import json
print('directed:', json.loads(open('graphify-out/graph.json').read())['directed'])
"
# → directed: False

The Python --cluster-only flow used in the skill doesn't have this issue — node_link_graph reads the directed flag, returns a DiGraph, and to_json/node_link_data preserves it. It looks like the CLI subcommand is loading or saving without preserving the directedness.

Why it matters

CLAUDE.md / skill.md make --directed a contract: "switching to undirected requires a full rebuild, not --update". Users who run graphify cluster-only . after a directed --update discover the contract was broken silently. Edge directions can't be recovered from the resulting undirected graph because NetworkX canonicalises them on export.

Recovery requires rebuilding from semantic cache with build_from_json(extraction, directed=True) and re-running any post-extraction merges.

Suggested fix

In whatever code path the CLI cluster-only subcommand uses to load/save graph.json, mirror the skill's pattern: read directed from the JSON, build a DiGraph accordingly, and pass through to to_json (which already preserves directedness for DiGraph input).

Workaround

Don't use graphify cluster-only via CLI on directed corpora. Use the Python flow (skill --cluster-only step, or local script invoking cluster() and to_json() directly).