fix: smaller per-device batch + grad accum + bigger completion budget for L40S

fix: pass --per-device-train-batch-size 4 to GRPO so effective batch divides num_generations

add: pre/post eval + summarize + bumped GRPO config (space/training/app.py)

add: pre/post eval + summarize + bumped GRPO config (training/summarize.py)

add: pre/post eval + summarize + bumped GRPO config (training/evaluate.py)

fix: multi-GPU SFT shape mismatch (space/training/app.py)

fix: multi-GPU SFT shape mismatch (training/sft_warmstart.py)

fix: refresh env._latent after step loop in oracle collector

fix: accept --difficulty in sft_warmstart to match space/training/app.py

fix: include requirements-train.txt + tests (glob bug)

fix: add top-level requirements-unsloth.txt referenced by space/training/requirements.txt

fix: add top-level requirements-train.txt referenced by space/training/requirements.txt

fix: add top-level requirements.txt referenced by space/training/requirements.txt

space: add root Dockerfile for trainer Space

initial: drugenv trainer control panel

initial commit