re-verify-logic（重新验证逻辑）

Logic Verification — Reverse Engineering Phase 2 Critic

Verify logic diagrams produced by re-visualize-logic against actual source code. Runs as an independent Critic in a separate agent context.

Gate rule: Phase 3 (re-extract-requirements) shall not proceed until verification produces a PASS or WARN verdict for all components.

Three Principles (Critic Perspective)

1. Code is Truth

• The source code is the sole authority. If the logic diagram says something the code does not confirm, the diagram is wrong.

2. Traceability to Line

• Every line number in the diagram must resolve to actual code. Unresolvable references are hallucinations.

3. Behavior over Intent

• Verify what the diagram claims against what the code shows. Do not interpret or justify discrepancies.

Execution

Step 1: Load Artifacts

Read docs/reverse/{analysis}/manifest.json.

Determine which components to verify:

• If component argument is provided, verify only that component
• If omitted, verify all entries in phase2.completed that have verification: null

For each component, read the logic diagram: docs/reverse/{analysis}/02-logic-{component}.md

Step 2: Line Number Verification

For each node in the Mermaid diagram and each row in the node-to-line mapping table:

1. Extract the cited line number
2. Read the source file at that line
3. Compare the code at that line with the claim

Classify:

• ✅ MATCH: Code at cited line matches the description
• ⚠️ OFFSET: Code found within ±5 lines
• ❌ MISMATCH: Code at cited line does not match
• 🚫 HALLUCINATION: Described behavior not found in source at any location

Step 3: Mermaid Syntax Validation

Check each Mermaid code block for:

• Valid node definitions and arrow syntax
• No prohibited characters (!=, >=, [], () in text)
• Parseable structure (matching brackets, valid subgraphs)

Step 4: Completeness Check

Compare the logic diagram against the source method:

• Count control flow branches in source (if/else, switch cases, loops)
• Count corresponding nodes in the diagram
• Flag missing branches

Check side effects documentation:

• Grep source for database operations, network calls, file I/O, logging
• Verify each is documented in the side effects summary

Step 5: Generate Verification Report

Write to docs/reverse/{analysis}/verification/v2-logic-{component}.md:

# Logic Diagram Verification: {component}

**Verification Date**: {YYYY-MM-DD}
**Target Document**: 02-logic-{component}.md
**Verdict**: {PASS / WARN / FAIL}

## Summary

| Check | Result | Details |
|:------|:-------|:--------|
| Line Number Accuracy | {✅/⚠️/❌} | {match}/{total} nodes ({%}) |
| Mermaid Syntax | {✅/⚠️/❌} | {error_count} errors |
| Branch Coverage | {✅/⚠️/❌} | {covered}/{total} branches ({%}) |
| Side Effects | {✅/⚠️/❌} | {documented}/{total} side effects ({%}) |

## Verdict

{PASS/WARN/FAIL}: {explanation}

## Details

### Line Number Issues
| Node | Claimed Line | Actual Line | Issue |
|:-----|:-------------|:------------|:------|

### Missing Branches
- {description of missing branch at file:line}

### Missing Side Effects
- {description of undocumented side effect at file:line}

## Recommendations

1. **[{severity}]** {fix description}

Step 6: Update Manifest

For each verified component:

• Update phase2.completed[].verification to "verification/v2-logic-{component}.md"

After all components are verified:

• If all verdicts are PASS or WARN: set phase2.status to "verified"
• If any verdict is FAIL: leave phase2.status as "completed"
• Update updated timestamp

Verdict Criteria

Any single FAIL criterion results in overall FAIL.

Prohibited Actions

• Do NOT modify logic diagram documents
• Do NOT modify source code files
• Do NOT approve documents with FAIL status
• Do NOT access the Doer's conversation context