[fix] train on trial-aligned windows for correct R² evaluation

[peterajhgraham]

Summary

• Replaces NLBDataset (sliding windows over full recording) with TrialAlignedDataset via build_trial_aligned_datasets in scripts/train_benchmark.py
• Sliding windows sample ~85% rest periods where velocity ≈ 0, producing R² ≈ 0 regardless of model quality; trial-aligned windows match the published NLB evaluation protocol
• Each sample is now a fixed [-100ms, +500ms] window around move_onset_time, yielding meaningful R² (Wiener filter baseline ~0.49)
• Replaces --data-root CLI arg with --nwb-path pointing directly to the MC_Maze train+behavior NWB file
• Removes the now-unnecessary val subsetting (trial datasets are already small — one sample per trial)
• Renames train_windows/val_windows keys in the JSON output to train_trials/val_trials

Test plan

• Run python scripts/train_benchmark.py --nwb-path <path-to-nwb> --max-steps 100 and confirm it loads without error
• Confirm train_trials and val_trials appear in benchmarks/training/results_raw.json
• Confirm R² after convergence is meaningfully above 0 (expected 0.3–0.6 range)

https://claude.ai/code/session_01AcwBjnbswWKkeyujGsuFzy

---

Generated by Claude Code