feat: switch default model to MiniMax M2.5

[LearningCircuit]

Summary

• Switch default AI model from moonshotai/kimi-k2-thinking to minimax/minimax-m2.5
• Updated in both ai-reviewer.sh and .github/workflows/ai-code-reviewer.yml

Why MiniMax M2.5?

• 80.2% on SWE-Bench Verified - excellent for code review tasks
• 204,800 context tokens - larger context window for bigger PRs
• 131,072 max output tokens - comprehensive reviews
• Reasoning support - compatible with existing <thinking> tag handling
• Cost effective - $0.30/M input, $1.20/M output via OpenRouter

Test plan

• Add ai_code_review label to a test PR
• Verify the AI review is generated using MiniMax M2.5
• Confirm review quality and response format are correct

[github-actions]

AI Code Review

Summary: Simple, clean change switching the default AI model. No issues found.

🔒 Security

• No security concerns identified. Model name change is a safe string substitution with no exposure of secrets or sensitive data.

Code Quality & Best Practices

• ✅ Consistent change across both config files (workflow + shell script)
• ✅ Proper use of bash default syntax ${VAR:-default} in ai-reviewer.sh
• ✅ Proper use of YAML default syntax ${{ vars.AI_MODEL || 'default' }} in workflow
• ✅ Backward compatible - users can override via AI_MODEL environment variable

Notes

• Model choice rationale is well-documented in PR (SWE-Bench score, context window, cost)
• The MAX_DIFF_SIZE differs between files (5000000 in shell vs 800000 in workflow) - verify this intentional discrepancy if relevant

---

✅ Approved

Review by Friendly AI Reviewer - made with ❤️