cluster-only on a merge-graphs output remaps against per-source community ids, so saved LLM labels are mostly lost on every rebuild #3858

Summary

When cluster-only re-clusters, it maps the new communities onto the previous ones
(remap_communities_to_previous, #822/#1027) using the community attribute of the nodes in
the input graph.json (cli.py, previous_node_community = {n["id"]: n["community"] ...}).

In a multi-source pipeline, that input is the fresh output of merge-graphs, which carries
each source's own community ids (overlapping 0..N per source before 0.9.50, offset per
input since 0.9.50, original kept in local_community). It does not carry the previous
merged clustering, which is what .graphify_labels.json / .graphify_labels.json.sig
were written against. The remap therefore aligns new communities to the wrong groups, the
per-cid signature check fails for most communities, and they are renamed by their hub.

Version

Measured on 0.9.46; code path unchanged in 0.9.69 (label-reuse branch identical by diff,
previous_node_community still read from the input file).

Evidence

Pipeline: per-source graphify update → graphify merge-graphs a.json b.json ... --out merged/graphify-out/graph.json → cd merged && graphify cluster-only . (labels file kept).
~15 000 nodes, ~1 700 communities, 5 sources.

The saved merged graph.json from the day before had 660 distinct community values
after merge-graphs (per-source ids) vs 1 709 in the previous merged clustering.

Offline simulation on the same graph (0.9.69 offsets, no workaround): if one early source
gains a single community, all later offsets shift and only 285 / 1 702 labels survive.

Workaround (10 lines)

cp merged/graphify-out/graph.json merged/graphify-out/graph.previous.json
graphify merge-graphs ... --out merged/graphify-out/graph.json
python3 - <<'PY'
import json
prev = {n["id"]: n["community"] for n in json.load(open("merged/graphify-out/graph.previous.json"))["nodes"]
        if n.get("community") is not None}
g = json.load(open("merged/graphify-out/graph.json"))
for n in g["nodes"]:
    if n["id"] in prev: n["community"] = prev[n["id"]]
    else: n.pop("community", None)
json.dump(g, open("merged/graphify-out/graph.json", "w"))
PY
(cd merged && graphify cluster-only .)

Related

#3014 (merge-graphs namespacing — this is the downstream effect on label reuse), #2452 (seed
cluster() with a previous partition), #2298 (exact-membership signature), #3334.

Suggested fix

Either merge-graphs could accept --previous <merged graph.json> and emit the previous merged
community (keeping per-source ids only in local_community), or cluster-only could prefer
graphify-out/graph.json already on disk (the previous run's output) over the input's
community attribute when a labels sidecar exists.