Follow-ups to #1080: batch subject removal, self-loop deletion, test hardening

[alistair3149]

Addresses the review findings on #1080. Targets the PR branch (fix/544-referenced-subject-stubs), not master, so it can be folded into that PR before it merges.

What each commit does

Batch subject removal — src/GraphDatabasePlugins/Neo4j/Persistence/Neo4jProjectionStore.php

• removeAbsentSubjects and deletePage now route through one removeSubjects(), which classifies referenced-vs-unreferenced subjects in a single query and deletes the unreferenced ones in a single query, replacing the per-subject round-trip loop. Both call sites stay unified on the same stub/delete logic (the point of the fix in Keep referenced Subjects as stubs instead of deleting them #1080), so this does not re-introduce a savePage/deletePage split.
• The classification excludes self-loops (other <> subject), so a subject whose only incoming relation is its own self-reference is deleted rather than left as an unreachable orphan stub.
• reduceSubjectToStub gets WITH DISTINCT subject so the property/label reset runs once, not once per outgoing relation (the two OPTIONAL MATCHes form a cartesian product).
• getSubjectIdsByPageId stops returning the node properties and label list that both callers discarded.

Test hardening — tests/…/Neo4jProjectionStoreTest.php

• testFlippingASubjectBetweenMainAndChildLeavesASingleHasSubjectEdge locks detachSubjectsFromPage, which had no coverage. Because HasSubject is MERGEd with its isMain property, flipping a persisting subject's role would leave a duplicate edge without that method. Teeth-verified: the test goes red when the detachSubjectsFromPage call is removed.
• testRemovingASelfReferencingSubjectDeletesItRatherThanStubbingIt — verified red on the pre-fix code (a self-loop subject was kept as a stub), green after.
• assertRelationExists optionally asserts the preserved relation's id, so a regression that keeps an incoming relation by endpoints and type but drops its id is caught. Pinned at the four preservation call sites.

Behavioural note for the reviewer

Batching the classification changes one acknowledged-limitation edge: removing two mutually-referencing subjects in one save now leaves both as orphan stubs (deterministic) instead of one stub + one deleted (order-dependent). Both outcomes are within the acknowledged, harmless, self-healing orphan-stub limitation from #1080; the proper cascade-sweep fix is deferred and tracked in #1112. Flagging it because it is a minor semantic shift.

Validation

Run locally against the dev stack (@group Database):

• Neo4jProjectionStore 34 ✓, Neo4jSubjectRelationUpdater 8 ✓
• Projection consumers: ReplaceSubjectApi ✓, Neo4jPageIdentifiersLookup ✓, Neo4jSubjectLabelLookup ✓, MediaWikiPageDeleter ✓
• make cs (phpcs 579 files + phpstan) ✓
• Live smoke on the dev wiki: deletePage on a referenced subject reduced it to a correct stub with its incoming relations preserved; a full graph rebuild restored the projection to baseline.

Draft because the full suite was not run locally; CI is the gate for the remainder.

_{AI-authored — Claude Code, Opus 4.8 (1M context), as follow-ups to a code review of #1080. Each production change validated against the touched + consumer @group Database tests locally, plus a live smoke on the dev wiki; not yet human-reviewed.}