func TestRetrievalRecallGate(t *testing.T)

	// Floors. Each gold query's relevant candidates are lexically present AND semantically near the query, and
	// the org holds only twelve candidates, so a correct hybrid search ranks the relevant ones at or near the
	// top and recall at ten is essentially complete. The floors sit well below the expected healthy values so
	// a passing run is unambiguous, while a real regression (embeddings drift, HNSW ef_search cut, a broken
	// mission gate, or an RRF k constant change that scrambles the fusion) drops a metric below its floor.
	//
	// meanRecall@10 >= 0.8: a healthy search recalls nearly every relevant id within the top ten of a twelve
	// candidate pool, so 0.8 is a conservative floor that a correct search clears comfortably and a recall
	// regression breaks.
	if meanRecall10 < 0.8 {
		t.Fatalf("REGRESSION: mean recall@10 = %.3f is below the 0.8 floor; the hybrid search is no longer recalling the relevant candidates", meanRecall10)
	}
	// meanMRR >= 0.5: a healthy search puts a relevant candidate in the top two on average (rank one is 1.0,
	// rank two is 0.5), so 0.5 is a conservative floor on first relevant rank that a ranking regression breaks.
	if meanMRR < 0.5 {
		t.Fatalf("REGRESSION: mean MRR = %.3f is below the 0.5 floor; the first relevant candidate is ranking too deep", meanMRR)
	}

meanRecall@10 >= 0.8: a healthy search recalls nearly every relevant id within the top ten of a twelve