tests/verify_swiss_watch.py

import json
from pedagogical_builder import merge_and_verify_explanations, LLMSchemaError

def test_golden_merge():
    print("🚀 Running The Golden Test: V2.5.3 Swiss Watch")

    # 1. Mock SymPy Truth Nodes (Immutable)
    sympy_nodes = [
        {"step_id": 1, "math": "f(x) = x^2", "logic": "מבנה הפונקציה"},
        {"step_id": 2, "math": "f'(x) = 2x", "logic": "גזירה"}
    ]

    # 2. Mock LLM Pedagogical Explanations (Skin)
    llm_explanations = [
        {"step_id": 1, "pedagogical_explanation": "נתונה לנו פונקציה ריבועית פשוטה."},
        {"step_id": 2, "pedagogical_explanation": "נשתמש בכלל החזקה כדי למצוא את הנגזרת."}
    ]

    try:
        # 3. Perform the Merge
        final_nodes = merge_and_verify_explanations(sympy_nodes, llm_explanations)
        
        print("  ✅ Merge Success!")
        assert len(final_nodes) == 2
        assert final_nodes[1]["math"] == "f'(x) = 2x"
        assert final_nodes[1]["pedagogical_explanation"] == "נשתמש בכלל החזקה כדי למצוא את הנגזרת."
        
    except LLMSchemaError as e:
        print(f"  ❌ Merge Failed: {e}")
        assert False, "Golden Merge should have succeeded"

    # 4. Test Schema Violations
    print("\n🛡️ Testing Schema Enforcement (Swiss Watch Locks):")
    
    # Violation A: Length Mismatch
    try:
        merge_and_verify_explanations(sympy_nodes, llm_explanations[:1])
        assert False, "Should have failed on length mismatch"
    except LLMSchemaError:
        print("  ✅ Length Mismatch Guard - PASSED")

    # Violation B: Unexpected Keys (No Math allowed from LLM)
    llm_with_math = [
        {"step_id": 1, "pedagogical_explanation": "...", "math": "hacked math"},
        {"step_id": 2, "pedagogical_explanation": "..."}
    ]
    try:
        merge_and_verify_explanations(sympy_nodes, llm_with_math)
        assert False, "Should have failed on unexpected keys (math)"
    except LLMSchemaError:
        print("  ✅ Key Enforcement Guard - PASSED")

    # Violation C: Step Order Modified
    llm_shuffled = [
        {"step_id": 2, "pedagogical_explanation": "Step 2"},
        {"step_id": 1, "pedagogical_explanation": "Step 1"}
    ]
    try:
        merge_and_verify_explanations(sympy_nodes, llm_shuffled)
        assert False, "Should have failed on step order"
    except LLMSchemaError:
        print("  ✅ Step Order Guard - PASSED")

    print("\n🏆 The Golden Test: V2.5.3 Swiss Watch - VERIFIED!")

if __name__ == "__main__":
    test_golden_merge()