feat: add memory management pipeline

Admission control, entropy-based micro-faulting, phantom tool
injection for backing store queries, and xMemory session hierarchy
for long conversations (50+ turns).

Ultraworked with [Sisyphus](https://github.com/code-yeongyu/oh-my-opencode)

Co-authored-by: Sisyphus <clio-agent@sisyphuslabs.ai>

tests/test_entropy.py

@@ -0,0 +1,182 @@
+"""Tests for the entropy-gated faulting module.
+
+Tests the EntropyDetector's ability to detect hedging, uncertainty,
+evicted entity references, and the should_fault decision logic.
+"""
+
+from __future__ import annotations
+
+from mnemosyne.entropy import (
+    DEFAULT_FAULT_THRESHOLD,
+    EntropyDetector,
+    EntropySignal,
+)
+
+
+# ── Hedging detection ────────────────────────────────────────────────────
+
+
+def test_entropy_hedging_detected():
+    """Hedging language like 'I think' and 'probably' is detected."""
+    detector = EntropyDetector()
+    signal = detector.analyze_response(
+        "I think the function is probably in utils.py",
+        evicted_entities=[],
+    )
+    assert signal.has_hedging is True
+    assert signal.score > 0
+
+
+def test_entropy_multiple_patterns():
+    """Multiple hedging patterns increase the score."""
+    detector = EntropyDetector()
+    signal_one = detector.analyze_response(
+        "I think it might work.",
+        evicted_entities=[],
+    )
+    signal_many = detector.analyze_response(
+        "I think it probably might be in the file, if I recall correctly. I believe so.",
+        evicted_entities=[],
+    )
+    assert signal_many.score >= signal_one.score
+
+
+# ── Uncertainty detection ────────────────────────────────────────────────
+
+
+def test_entropy_uncertainty_detected():
+    """Explicit uncertainty like 'I'm not sure' is detected."""
+    detector = EntropyDetector()
+    signal = detector.analyze_response(
+        "I'm not sure about the exact implementation. It's unclear to me.",
+        evicted_entities=[],
+    )
+    assert signal.has_uncertainty is True
+    assert signal.score > 0
+
+
+# ── No signals ───────────────────────────────────────────────────────────
+
+
+def test_entropy_no_signals():
+    """Clean, confident response has no entropy signals."""
+    detector = EntropyDetector()
+    signal = detector.analyze_response(
+        "The function `calculate_total` is defined in src/utils.py at line 42. "
+        "It takes two arguments: price and quantity.",
+        evicted_entities=[],
+    )
+    assert signal.has_hedging is False
+    assert signal.has_uncertainty is False
+    assert signal.references_evicted is False
+    assert signal.has_hallucination_risk is False
+    assert signal.score == 0.0
+
+
+# ── Evicted entity references ───────────────────────────────────────────
+
+
+def test_entropy_evicted_references():
+    """References to evicted entities are detected."""
+    detector = EntropyDetector()
+    signal = detector.analyze_response(
+        "The config is in settings.py and the handler is in views.py",
+        evicted_entities=["settings.py", "views.py", "models.py"],
+    )
+    assert signal.references_evicted is True
+    assert "settings.py" in signal.referenced_entities
+    assert "views.py" in signal.referenced_entities
+    assert "models.py" not in signal.referenced_entities
+
+
+def test_entropy_hallucination_risk():
+    """Hedging + evicted references = hallucination risk."""
+    detector = EntropyDetector()
+    signal = detector.analyze_response(
+        "I think the config is probably in settings.py somewhere",
+        evicted_entities=["settings.py"],
+    )
+    assert signal.has_hedging is True
+    assert signal.references_evicted is True
+    assert signal.has_hallucination_risk is True
+    # Hallucination risk adds a bonus to the score
+    assert signal.score > 0.3
+
+
+# ── Composite score ──────────────────────────────────────────────────────
+
+
+def test_entropy_composite_score():
+    """Score combines hedging, uncertainty, and evicted references."""
+    detector = EntropyDetector()
+    # All signals present
+    signal = detector.analyze_response(
+        "I'm not sure, but I think the code is probably in utils.py",
+        evicted_entities=["utils.py"],
+    )
+    assert signal.has_hedging is True
+    assert signal.has_uncertainty is True
+    assert signal.references_evicted is True
+    assert signal.score > 0.5  # Should be high with all signals
+
+
+# ── should_fault decisions ───────────────────────────────────────────────
+
+
+def test_entropy_should_fault_above_threshold():
+    """Faulting triggers when score >= threshold AND references evicted."""
+    detector = EntropyDetector(fault_threshold=0.4)
+    signal = detector.analyze_response(
+        "I think the code is probably in utils.py, if I recall correctly. I believe so.",
+        evicted_entities=["utils.py"],
+    )
+    # Score should be above 0.4 with hedging + evicted ref + hallucination bonus
+    entities = detector.should_fault(signal)
+    assert len(entities) > 0
+    assert "utils.py" in entities
+
+
+def test_entropy_should_fault_below_threshold():
+    """No faulting when score is below threshold."""
+    detector = EntropyDetector(fault_threshold=0.99)  # Very high threshold
+    signal = detector.analyze_response(
+        "I think it might be there.",
+        evicted_entities=[],
+    )
+    entities = detector.should_fault(signal)
+    assert entities == []
+
+
+def test_entropy_should_fault_no_evicted_refs():
+    """No faulting when there are no evicted entity references, even with high hedging."""
+    detector = EntropyDetector(fault_threshold=0.1)  # Very low threshold
+    signal = detector.analyze_response(
+        "I think it probably might be somewhere, if I recall. I believe so.",
+        evicted_entities=[],
+    )
+    # Even though hedging score is high, no evicted refs → no fault
+    entities = detector.should_fault(signal)
+    assert entities == []
+
+
+# ── Stats tracking ───────────────────────────────────────────────────────
+
+
+def test_entropy_stats_tracking():
+    """Stats correctly count analyzed responses and triggered faults."""
+    detector = EntropyDetector(fault_threshold=0.3)
+    assert detector.stats == {"analyzed": 0, "faults_triggered": 0}
+
+    # Analyze a clean response
+    detector.analyze_response("The function is in utils.py.", evicted_entities=[])
+    assert detector.stats["analyzed"] == 1
+    assert detector.stats["faults_triggered"] == 0
+
+    # Analyze and fault a hedging response with evicted refs
+    signal = detector.analyze_response(
+        "I think the code is probably in utils.py, if I recall correctly. I believe so.",
+        evicted_entities=["utils.py"],
+    )
+    assert detector.stats["analyzed"] == 2
+    detector.should_fault(signal)
+    assert detector.stats["faults_triggered"] == 1