feat(plagiarism): показывать тематически близкие источники, а не только нарушения

Кандидаты уровня 3 (FAISS), которые прошли порог семантической схожести,
но LLM не подтвердила заимствование, раньше молча отбрасывались. Теперь
это отдельный блок "recommendations" в отчёте — не плагиат, но источники,
полезные для раскрытия темы.

services/worker-gpu/tests/test_scoring.py

@@ -139,3 +139,56 @@ def test_no_full_text_means_nothing_marked_cited():
     out = aggregate_results([_m("A", 0)], [], [], total_fragments=2)  # full_text не передан
     assert out["matches"][0]["cited"] is False
     assert out["overall_similarity"] == out["uncited_similarity"]
+
+
+# ─── aggregate_results: recommendations ──────────────────────────────────────
+
+def _cand(title: str, similarity: float, url: str | None = None) -> dict:
+    return {
+        "fragment": "фрагмент",
+        "position_start": 0,
+        "position_end": 8,
+        "similarity": similarity,
+        "source_title": title,
+        "source_url": url,
+        "source_db": "openalex",
+    }
+
+
+def test_no_related_candidates_means_empty_recommendations():
+    out = aggregate_results([], [], [], total_fragments=10)
+    assert out["recommendations"] == []
+
+
+def test_related_candidates_become_recommendations_sorted_by_similarity():
+    out = aggregate_results(
+        [], [], [], total_fragments=10,
+        related_candidates=[_cand("Low", 60.0), _cand("High", 90.0)],
+    )
+    titles = [r["source_title"] for r in out["recommendations"]]
+    assert titles == ["High", "Low"]
+
+
+def test_recommendations_dedup_keeps_best_per_source():
+    out = aggregate_results(
+        [], [], [], total_fragments=10,
+        related_candidates=[_cand("A", 60.0, "url-a"), _cand("A", 85.0, "url-a")],
+    )
+    assert len(out["recommendations"]) == 1
+    assert out["recommendations"][0]["similarity"] == 85.0
+
+
+def test_recommendations_capped_at_ten():
+    candidates = [_cand(f"S{i}", float(i), f"url-{i}") for i in range(15)]
+    out = aggregate_results([], [], [], total_fragments=10, related_candidates=candidates)
+    assert len(out["recommendations"]) == 10
+
+
+def test_source_already_flagged_as_match_excluded_from_recommendations():
+    semantic = [{**_cand("A", 90.0, "url-a"), "method": "semantic+llm"}]
+    out = aggregate_results(
+        [], [], semantic, total_fragments=10,
+        related_candidates=[_cand("A", 60.0, "url-a"), _cand("B", 70.0, "url-b")],
+    )
+    titles = [r["source_title"] for r in out["recommendations"]]
+    assert titles == ["B"]