fix: torch==2.5.1 (cu121 index only has up to 2.5.1)

fix: install torch with CUDA index URL, verify imports in build

Train deploy: configs/train/sft.yaml

Train deploy: configs/train/grpo.yaml

Train deploy: configs/reward/v1.yaml

Train deploy: configs/model/qwen_1_5b.yaml

Train deploy: configs/model/qwen_0_5b_smoke.yaml

Train deploy: configs/data/v1.yaml

Train deploy: configs/config.yaml

Train deploy: env/server/__init__.py

Train deploy: env/server/repl.py

Train deploy: env/server/llm_primitive.py

Train deploy: env/server/environment.py

Train deploy: env/server/app.py

Train deploy: env/data/phase0_fixture.jsonl

Train deploy: env/__init__.py

Train deploy: env/models.py

Train deploy: env/client.py

Train deploy: data/__init__.py

Train deploy: data/train.jsonl

Train deploy: data/sft_traces.jsonl

Train deploy: data/seeds.json

Train deploy: data/generate.py

Train deploy: data/eval.jsonl

Train deploy: rewards/__init__.py

Train deploy: rewards/token_budget.py

Train deploy: rewards/recursion_efficiency.py

Train deploy: rewards/format_gate.py

Train deploy: rewards/correctness.py

Train deploy: rewards/compose.py

Train deploy: train/__init__.py

Train deploy: train/smoke_test.py

Train deploy: train/sft.py

Train deploy: train/model_load.py

Train deploy: train/grpo.py

Train deploy: pyproject.toml

Train deploy: run_training.py

Train deploy: README.md

Train deploy: Dockerfile

initial commit