chore: codebase hygiene pass — untrack weights, migrate to logging, tidy comments

fix: eliminate infinite-loop risk in maze start/goal sampling

feat(round4): four-algorithm ablation — Dueling best at 84% Holdout

fix(demo): strengthen anti-loop by penalizing moves toward high-frequency cells

fix(demo): auto-infer input_channels from checkpoint weight shape

fix(demo): stop animation on any user interaction

fix(demo): re-enable inference-side anti-loop Q-penalty

feat(demo): Streamlit web demo — Plotly heatmap, anti-loop inference