Summary

• fix evaluation entry by switching src/browser_control/evaluate.py to a valid debug config
• align local GRPO training with the Android inference flow by updating prompts, action parsing, reward shaping, and adding configs/lfm2_350m_local_full_v2.yaml
• improve local Docker training by supporting an external MiniWoB server, adding AXTree debug injection, and stabilizing the training image setup
• update local GGUF conversion to use the latest checkpoint path and run without interactive confirmation
• normalize project entry docs with AGENT.md / README.md, remove legacy workflow files, refresh evaluation screenshots, and ignore workflow metadata directories in .gitignore

Testing

• verified local baseline workflow previously recorded in project notes:
  • uv sync
  • debug fine-tune run
  • make evaluation over 10 episodes
• refreshed media/episode_0_step_0.png through media/episode_9_step_0.png from the latest evaluation pass
• no additional full test/build run was executed after the final doc and ignore-rule cleanup

Notes

• .project/ and .learnings/ are intentionally excluded from version control
• the only remaining untracked item is the outer-repo file ../../docs/, which is not part of this project PR