tests/verify_v47_rollback.py

import json
import re
import logging
import asyncio
import sympy as sp
from strategy_manager import StrategyManager
from pedagogical_builder import build_pedagogical_response
from orchestrator import verify_math_consistency

# Configure logging
logging.basicConfig(level=logging.INFO)
logger = logging.getLogger(__name__)

async def test_v47_rollback():
    print("🚀 Starting V4.7 Rollback & SymPy Hardening Verification")
    
    # 1. Test SymPy Hardening (Risk: Equations crashing)
    print("\n⚖️ Testing SymPy Hardening:")
    
    # Test 1: Equation vs Value (Should be False for identity, even if consistent)
    anchor = "x^2 = 4"
    result = "2"
    is_id, scale = verify_math_consistency(anchor, result)
    print(f"  - Equation 'x^2 = 4' vs '2': Identical={is_id}, Scale={scale}")
    assert is_id == False, "Equation and value should not be identical"
    assert scale == 0.6, "Scale should be 0.6 for non-identity"

    # Test 1b: True Identity (Expression vs Expression)
    is_id, scale = verify_math_consistency("x^2 + 1", "1 + x^2")
    print(f"  - Identity 'x^2 + 1' vs '1 + x^2': Identical={is_id}, Scale={scale}")
    assert is_id == True, "Symbolic identity failed"
    assert scale == 1.0, "Scale should be 1.0 for identity"

    # Test 2: Invalid syntax handling
    is_id, scale = verify_math_consistency("!!!invalid!!!", "2")
    print(f"  - Invalid syntax handling: Identical={is_id}, Scale={scale}")
    assert is_id == False, "SymPy should return False on crash"

    # Test 3: Derivative Normalization (Hotfix #1)
    # Using explicit multiplication (*) as sympify is strict
    is_id, scale = verify_math_consistency("f'(x) = 2*x", "Derivative(f(x), x) = 2*x")
    print(f"  - Derivative 'f'(x)' normalization: Identical={is_id}, Scale={scale}")
    assert is_id == True, "Derivative normalization identity failed"

    # 2. Test Strategy Manager V4.7 Output
    sm = StrategyManager(None)
    raw_response = """
    {
      "confidence_score": 0.9,
      "is_valid": true,
      "latex_result": "x=2",
      "solution_markdown": "התשובה היא $x=2$."
    }
    """
    parsed = sm._parse_v4_json(raw_response)
    print("\n✅ Strategy Manager V4.7 Parsing: Passed")

    # 3. Test Pedagogical Builder Priority Bypass (Hotfix #2)
    ped_response = build_pedagogical_response(
        topic_id="GENERAL", # Template should be bypassed
        llm_output=parsed,
        data_anchor={}
    )
    steps = ped_response['sections'][0]['steps']
    print(f"\n📊 Pedagogical steps count (Priority Bypass): {len(steps)}")
    # Bypass logic returns precisely 1 step in 1 section
    assert len(steps) == 1, f"Priority bypass failed, got {len(steps)} steps"
    assert "התשובה היא" in steps[0]['content_mixed'], "Solution content mismatch"
    print("✅ Pedagogical Priority Bypass - PASSED")

    print("\n🏆 V4.7 Rollback & SymPy Hardening Verified!")

if __name__ == "__main__":
    asyncio.run(test_v47_rollback())