test: add T=1/T=2 integration smoke tests against real ZAYA1-8B Q4

tests/test_integration_smoke.py

@@ -0,0 +1,39 @@
+import pytest
+
+
+@pytest.mark.integration
+def test_t1_n2_real_zaya_smoke(tmp_path):
+    """T=1 N=2 on a tiny prompt — should produce non-empty text + complete audit."""
+    from markovian_rsa_mlx import MarkovianRSAOrchestrator, RSAConfig
+
+    orch = MarkovianRSAOrchestrator.from_pretrained("kyr0/zaya1-base-8b-MLX")
+    cfg = RSAConfig(
+        rounds=1, parallel=2, aggregation_subsample=2,
+        chunk_tokens=128, tail_tokens=32, serial=True, seed=0,
+    )
+    audit_path = tmp_path / "audit.jsonl"
+    text, result = orch.solve("What is the capital of France?",
+                              config=cfg, return_audit=True, audit_path=audit_path)
+    assert isinstance(text, str)
+    assert len(text) > 0
+    assert audit_path.exists()
+    audit_lines = audit_path.read_text().strip().split("\n")
+    assert len(audit_lines) >= 5
+    assert result.stats.total_generated_tokens > 0
+
+
+@pytest.mark.integration
+def test_t2_n2_real_zaya_smoke(tmp_path):
+    """T=2 N=2 sanity — same prompt should produce final text after aggregation."""
+    from markovian_rsa_mlx import MarkovianRSAOrchestrator, RSAConfig
+
+    orch = MarkovianRSAOrchestrator.from_pretrained("kyr0/zaya1-base-8b-MLX")
+    cfg = RSAConfig(
+        rounds=2, parallel=2, aggregation_subsample=2,
+        chunk_tokens=128, tail_tokens=32, serial=True, seed=0,
+    )
+    text, result = orch.solve("Compute 7*8.", config=cfg, return_audit=True,
+                              audit_path=tmp_path / "audit.jsonl")
+    assert len(result.rounds) == 2
+    assert all(len(r.traces) == 2 for r in result.rounds)
+    assert all(t.parent_trace_ids for t in result.rounds[1].traces)