docs: drop R4-续 retrospective paragraph — it reintroduced the wrong shaping=0.5 value

docs: polish pass — README wording, experiment_log structure + honest retrospective, LaTeX micro-style

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

docs: clean up R3/R4 record and consolidate technical narrative

docs: finalize R4 documentation — Dueling 84% Holdout, full ablation record

refactor(model): update architecture docs and set dueling as default algorithm

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

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

docs(round4): finalize R4 Double DQN results — 78% Holdout, Grid-SPL clarification

chore: add MIT LICENSE, ignore .claude/ directory

chore(weights): upgrade vanilla/dueling/double_dueling to 4-channel R4 weights

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

chore(weights): update double DQN weight to R4-A3 4-channel (78% holdout)

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

add model weights via Git LFS, fix HF Space build

docs: README — results table, architecture, quickstart, references

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

docs(round4): complete experiment record — A1/A2/A3 full EVAL data and conclusions

style(train): remove forward-reference quotes from type hints (Python 3.10+)

fix(train): use terminated-only mask for TD bootstrap (Gymnasium v0.26)

fix(train): guarantee BFS-connected start/goal, bounded retry with fallback

feat(round4): upgrade obs 3->4 channels (visited_map) + EVAL-based checkpoint

feat(round3): buffer=80k + target_freq=1500 + shaping=0.5 → 74% holdout, SPL=0.735

feat(round2): extended training, Double DQN 64% holdout, SPL=0.633

feat(round1): baseline DQN variants — Vanilla/Double/Dueling/Double+Dueling

ci: GitHub Actions — pytest + 90%+ branch coverage gate

feat(env): Gymnasium maze env, 3-channel obs, BFS reachability

chore: initial project scaffold