Add SFT warm start before GRPO and DB connectivity init check

Add local model inference backend for Layer 2

Increase max completion length from 512 to 2048

Add Supabase upload for training results (Storage + DB)

Improve reward function to break refuse-everything local minimum and scale training

Clean up dead code, unused imports, and move hardcoded values to config.yaml

Centralize all training params in config.yaml (single source of truth)