Fix num_generations=2

SFT 200->350 steps for better format compliance

Fix curriculum window=50 and threshold=0.80

GRPO speed fix: 50 steps, 1 generation

Add WHY comments + GRPO/SFT speed caps

Speed fix: cap SFT at 200 steps

Update train_data_centric.py

Audit fixes: remove duplicate torch import, add metadata field, fix stale strings, fix test assertions, update reward docs

Redesign reward for discrimination: efficiency multiplier, strict penalties, stretch bonus, start at level 1

Fix demo mode crash: use max_steps param instead of unpicklable local class

Optimize for fast iteration: 1.5B model, LoRA r=8, GRPO batch=2/gen=2, seq=512

Switch experiment tracking from W&B to TensorBoard (no API key required)

Enable W&B experiment tracking in SFT+GRPO phases (required by hackathon)

refactor: extract agent_utils.py (shared prompt/commands/server utils), simplify reward to env+format, add audit.py

Data-Centric AI RL Environment — OpenEnv Hackathon Submission