sft+reward-fix: space/training/app.py

sft+reward-fix: training/training_script.py

sft+reward-fix: training/sft_warmstart.py

sft+reward-fix: tests/test_reward_hacking.py

sft+reward-fix: server/environment.py

sft+reward-fix: server/rewards/reward_function.py

fix: coerce beam_energy to str so CollisionObservation pydantic check accepts numeric LLM outputs

vanilla GRPO: backport EvidenceCallback for live evidence/*.csv + plots

dashboard: synthesize PNGs on demand + cache-bust + pass --evidence_dir to vanilla

fix: peft 0.18 -> 0.13.2 to match transformers 4.51.3 (vanilla path)

fix(deps): peft 0.18.0 -> 0.13.2 to match transformers 4.51.3

fix: switch trainer Space to vanilla GRPO path

fix: switch trainer Space to vanilla GRPO path

fix: disable fast_inference (vLLM not installed) in training/evaluate.py

fix: disable fast_inference (vLLM not installed) in training/training_unsloth.py

Update CERNenv Space

initial commit