Delete _patch_diagnostics.py

Fix: backward() inside @torch .no_grad() — use torch.enable_grad() for dense gradient computation

Fix: compute_relaxer_diagnostics called backward inside no_grad context"

Upload exp5_mechanism.py with huggingface_hub

Upload exp4_relaxation.py with huggingface_hub

Upload results/exp3.json with huggingface_hub

Upload results/exp2.json with huggingface_hub

Upload results/exp1.json with huggingface_hub

Upload ablations_lite.py with huggingface_hub

Major revision: add phantom momentum ablation, compute-matched baselines, multi-seed predictor accuracy

Upload ablations.py with huggingface_hub

Add definitive ablation suite addressing all critique gaps

Add sparse transformer v19 with Triton-backed KNN scheduler and various backward modes. Includes utilities for synthetic data generation and model training. Implements chunked sparse updates and integrates with existing sparse linear layers.

Upload triton_v2.py with huggingface_hub

Add complete results with all measured numbers

Upload e2e_bench.py with huggingface_hub

Upload e2e_full.py with huggingface_hub

Upload triton_sparse.py with huggingface_hub

Add all experiment code

Upload README.md

initial commit