Orphan stubs when removing mutually-referencing subjects in one save (cascade sweep)

[alistair3149]

Problem

When a Subject is removed from its page (or its page is deleted) while still referenced by another Subject, the Neo4j projection reduces it to a stub rather than deleting it, keeping the incoming relation valid (#1080). But when two or more subjects that reference each other are removed in the same savePage/deletePage, the removals cascade: stubbing one drops the relation that made another "referenced", so a subject can end up as a stub with no remaining incoming reference — an orphan stub.

Example (A -[rel]-> B and B -[rel]-> A, both removed in one save):

• Both are classified as referenced (each has an incoming relation from the other).
• Reducing each to a stub deletes its outgoing relations, so after both are processed each has lost the relation that referenced it.
• Result: two orphan stub nodes — :Subject nodes with only id + wiki_id, no name, no schema label, no HasSubject edge, and no incoming relation.

Path: Neo4jProjectionStore::removeSubjects() / reduceSubjectToStub().

Impact

Harmless in the supported readers, and self-healing:

• Neo4jSubjectLabelLookup filters on name + schema label — a stub has neither, so it never surfaces.
• Neo4jPageIdentifiersLookup traverses :HasSubject — a stub has none.
• RebuildGraphDatabases.php rebuilds the projection from scratch and normalises them away.

But a raw MATCH (n:Subject) via nw.query (Lua), the Cypher REST API, or {{#cypher_raw}} does enumerate orphan stubs (they keep the :Subject label). No stale data leaks — a stub carries only id + wiki_id — so this is phantom-node / count-inflation, not data exposure.

Robust fix

Cascade-sweep newly-orphaned stubs at the end of the write transaction: after the removals, delete any :Subject node left with no HasSubject edge and no incoming non-HasSubject relation (i.e. unreachable).

Notes / history

• Acknowledged as a deferred limitation in Keep referenced Subjects as stubs instead of deleting them #1080.
• The self-loop degenerate case (a subject referencing itself) is already fixed in Follow-ups to #1080: batch subject removal, self-loop deletion, test hardening #1111: self-loops are excluded from the incoming-relation check, so a self-referencing subject is deleted rather than stubbed. This issue is specifically about multi-subject reference cycles.
• Follow-ups to #1080: batch subject removal, self-loop deletion, test hardening #1111 also made the multi-subject case deterministic (batched classification → both subjects become orphan stubs) rather than order-dependent (one stub + one deleted on master). Same harmless class either way.

Low priority: NeoWiki is not in production, the graph is a rebuildable projection, and the stubs are invisible to the structured readers.