Add Gemini 3 Pro to bash-only leaderboard (#53)

* Add GPT 5.2 Codex (high reasoning) and fix GPT 5.2 naming on leaderboards

* Add Gemini 3 Pro to bash-only leaderboard

Resolved: 69.6% (348/500), Cost: $480.01, mini-swe-agent v2.0.0

* Fix os_model flags for open-weights models in leaderboard data

Fix os_model: false → true for DeepSeek V3.2, GLM-5, Kimi K2.5, and
MiniMax M2.5 (high reasoning) in both bash-only and verified leaderboards.

* Fix Gemini 3 Pro resolved counts in leaderboard

Regenerated leaderboards.json with corrected per_instance_details
showing 348/500 resolved (69.6%) instead of 0.

Author: aorwall