04 — Force graphify overwrite after page removal

Priority: P1
Depends on: none
Size: S (15–30 min)
Files: sessionwiki/graphify_bridge.py, tests/test_graphify_bridge.py
PR title: fix: force graphify overwrite after page removal

Read first

sessionwiki/graphify_bridge.py, function build_graphify_graph (around line 221), especially the to_json(...) call near line 332
tests/test_graphify_bridge.py
graphify/export.py, signature of to_json (installed package): it accepts force: bool = False

Problem

After pages are deleted or coalesced, sessionwiki graph prints:

[graphify] WARNING: Refusing to overwrite — saved semantic node(s) are missing

Then it falls back to the builtin engine and leaves graphify-out/ stale. The cause: to_json protects an existing graph from shrinking. sessionwiki regenerates the graph from wiki/, so the old graph should be replaced.

This happened after the 54-to-1 session coalescing. The workaround was rm -rf graphify-out, which is not acceptable for users.

Required change

In sessionwiki/graphify_bridge.py change:

to_json(G, communities, str(json_path))

to:

to_json(G, communities, str(json_path), force=True)

Add a short comment:

# sessionwiki owns this graph: it is regenerated from wiki/ on every run.
# Without force=True, graphify refuses to shrink a graph after pages
# are deleted or coalesced.

Do not pass force=True to to_html or other exporters unless their signatures accept it.

Test

build_graphify_graph is heavy and depends on the optional graphify package. A unit test that runs the whole pipeline is impractical.

Add one focused test to tests/test_graphify_bridge.py that checks the call contract. Keep it skipped when graphify is not installed:

@pytest.mark.skipif(
    not _GRAPHIFY_INSTALLED,
    reason="graphify is an optional dependency",
)
def test_build_graphify_graph_forces_overwrite():
    import inspect
    from sessionwiki.graphify_bridge import build_graphify_graph

    source = inspect.getsource(build_graphify_graph)
    assert "force=True" in source
    assert "to_json(" in source

The source assertion is intentionally narrow. It guards the regression without building a graph.

Integration check

This is the real acceptance test. Use a disposable source page so no content is lost:

cp wiki/sources/llm-wiki/2026-10-29-session_d45df3ef6596bc2f63c549dda7fc2060.md /tmp/keep.md
mv wiki/sources/llm-wiki/2026-10-29-session_d45df3ef6596bc2f63c549dda7fc2060.md /tmp/deleted.md
.venv/bin/python -m sessionwiki graph 2>&1 | tee /tmp/graph-check.log
! grep -q "Refusing to overwrite" /tmp/graph-check.log

Expected: graphify writes graphify-out/graph.json and prints no "Refusing to overwrite" warning.

Restore the page and rebuild:

mv /tmp/deleted.md wiki/sources/llm-wiki/2026-10-29-session_d45df3ef6596bc2f63c549dda7fc2060.md
.venv/bin/python -m sessionwiki graph

Check that the graph contains the expected node count:

python3 -c "import json; d=json.load(open('graphify-out/graph.json')); print(len(d['nodes']))"

Acceptance criteria

[ ] to_json(..., force=True) is in the bridge.
[ ] The focused test passes.
[ ] The integration check prints no "Refusing to overwrite".
[ ] graphify-out/graph.json is rewritten (mtime changes).
[ ] sessionwiki lint still reports 0 errors.
[ ] CHANGELOG.md updated.

Out of scope

Adding a --force CLI flag. Always forcing is correct here because sessionwiki owns the graph.
Changing graphify itself.
Deleting graphify-out in code.

Rollback

Remove force=True. Users can still recover with rm -rf graphify-out.