Add TODO.md — next steps for NSGF++ reproduction

Add LEARNING.md — all mistakes, war stories, examples, and principles from NSGF++ reproduction

Split SKILL.md into SKILL.md (rules) + LEARNING.md (stories/mistakes) + TODO.md (next steps)"

SKILL.md: Add VRAM estimation, checkpointing, multi-session, multi-GPU lessons from CIFAR OOM

config.yaml: CIFAR sinkhorn batch 128→32 for T4, pool batches 2500→10000 to compensate

main.py: --resume-phase, --checkpoint-dir, --sinkhorn-batch flags

Add checkpointing, resume, CIFAR OOM fix, --sinkhorn-batch flag

Update SKILL.md with DataLoader batch size mismatch bug and shared state pitfall

--train-iters now also overrides time_predictor.num_iterations

Fix DataLoader batch size mismatch across training phases + --train-iters now overrides all phases

Add pool.finalize() call after building trajectory pool for O(1) sampling

Fix geomloss tensor shape bug for images + optimize pool sampling

Add SKILL.md — paper reproduction skill with lessons from NSGF++ implementation

Create skill-creator.md

Upload main.py

Upload requirements.txt

Upload main.py

Upload evaluation.py

Upload inference.py

Upload trainer.py

Upload model.py

Upload sinkhorn_flow.py

Upload dataset_loader.py

Upload config.yaml

Upload README.md

initial commit