Improve system prompt to reduce false positives

[felipefernandes]

🎯 Goal

Reduce false positives in code reviews by improving the system prompt with specific guidelines.

📊 Complexity

Quick Win (1-2 hours)

🔍 Problem

The LLM is generating false positives like:

• Reporting ${{ secrets.X }} as hardcoded secrets (it's correct GitHub Actions syntax)
• Flagging os.chmod on config files as performance issues (it's security best practice)
• Reporting existing try-except blocks as "missing error handling"
• Suggesting premature optimizations for small lists (< 10 items)

✅ Solution

Add specific guidelines to the system prompt in iara/prompt.py:

IMPORTANT GUIDELINES:
- GitHub Actions: ${{ secrets.X }} is CORRECT, NOT hardcoded
- os.chmod on config files: Security best practice, NOT performance issue
- Existing try-except blocks: NOT missing error handling
- Small lists (< 10 items): O(n) vs O(1) is negligible, NOT worth optimizing
- Only report REAL problems that would cause bugs, security issues, or significant performance degradation
- When in doubt, DON'T report

📝 Implementation Steps

1. Read current system prompt in iara/prompt.py
2. Add new section with false positive guidelines
3. Test with known false positive cases
4. Verify reduction in false positive rate

🎁 Expected Impact

• 50-70% reduction in false positives
• More focused and actionable reviews
• Better user trust in the tool

---

Related to Groq provider integration (#67) where false positives were identified.