Fix WebSocket play/training paths + live-play gallery + gitignore hygiene

[aifriend]

Five commits picking up where PR #1 left off — three orthogonal improvements that came out of actually running the console end-to-end against the bundled checkpoints.

Summary

Fix: make WebSocket play + TRAIN 5K actually work (d2f7a2b)

Three real bugs that blocked the demo path the UI advertises:

1. NumPy 2.0 removed np.bool8, but gym 0.25.2 (pinned here) still references it in passive_env_checker.py during env.step. Restored the alias at the top of deep_rl_zoo/gym_env.py so every entry point (run_atari / run_classic / eval_agent / frontend.stream_eval) benefits. Defensive shim kept at the top of stream_eval.py for any future caller that imports gym before deep_rl_zoo.
2. PER-DQN checkpoint mismatch: frontend/stream_eval.py loaded PER-DQN_Pong_4.ckpt with agent_name="DQN", but the file is stamped "PER-DQN" (matches deep_rl_zoo/prioritized_dqn/run_atari.py:96). ▶ Play silently failed for PER-DQN. Fixed.
3. No DQN factory in ALGO_FACTORIES — any DQN_<game>.ckpt (including newly TRAIN-spawned ones) couldn't be replayed. Added _build_dqn.

Verified end-to-end by capturing all 4 bundled agents playing live and then training + replaying a fresh DQN/Breakout checkpoint via the same /api/training/start → /api/eval/stream path the UI uses.

Demo: live-play gallery (9bb6cb9)

screenshots/live_play/ now contains a self-contained gallery of real ALE frames from the live console paths:

• 5 trained agents (4 bundled + 1 freshly TRAIN-spawned DQN/Breakout) captured via the WebSocket.
• 6 random-policy renders across visually distinct games: SpaceInvaders, MsPacman, Asteroids, Q*bert, Seaquest, Boxing — captured directly via deep_rl_zoo.gym_env.create_atari_environment.

Each game has both a 5-frame montage PNG (static, with score progression) and an animated 100–120 frame GIF (motion). gallery.html embeds all of them in a single phosphor-styled page — open in any browser to see every agent and game in one place.

Captures use the same env.unwrapped.render(mode="rgb_array") path the WebSocket eval emits, so what's shown is frame-accurate to what the production UI renders.

Chore: clean git status for everyone (1ff25e5, a6e11cb, c86edda)

Every push and PR was leaving behind the same handful of untracked / modified files, polluting reviews and risking accidental staging. Root cause was an incomplete .gitignore plus a tracked file that the training subprocess auto-appends to.

Now ignored:

• .claude/ — Claude Code worktree config (per-session)
• .cursorindexingignore — Cursor IDE metadata
• runs/_training_*.log — per-job stdout from frontend/server.py
• .venv/ → .venv so the bare-name pattern catches both real directories and symlinks to a shared parent venv

Untracked (files stay on disk locally, historical snapshots remain in git history):

• logs/*_atari_results.csv — all 6 per-algorithm result CSVs that deep_rl_zoo.<algo>.run_atari appends to on every iteration. The dqn one was already dirtying the tree after this session's TRAIN; the other 5 (iqn, per_dqn, ppo_rnd, r2d2, rainbow) would have done the same the moment those algos got trained. Collapsed into a glob so any future algorithm's results file is covered with no further work.

Verified: git check-ignore confirms every previously-leaking path now resolves to an entry in .gitignore, and the working tree is clean after a full TRAIN run.

Changes

File group	What
deep_rl_zoo/gym_env.py	NumPy bool8 shim
frontend/stream_eval.py	Defensive shim + PER-DQN agent_name fix + new DQN factory
screenshots/live_play/	21 new files — montages, GIFs, gallery.html, summary.txt
.gitignore	+13 / −1 — covers leakers and per-algo result CSVs
logs/*_atari_results.csv	6 files removed from tracking (kept on disk locally)

Test plan

• ./start.sh, then click ▶ Play with each bundled checkpoint (Rainbow/Pong, PER-DQN/Pong, IQN/Pong, PPO-RND/MontezumaRevenge) — all four should stream frames; previously IQN/Rainbow/PPO-RND failed on np.bool8 and PER-DQN failed on the agent_name mismatch
• ▶ TRAIN 5K on DQN/Pong (or any other algo) — the subprocess should complete with returncode 0 and write a checkpoint; previously it failed on np.bool8 at train_env.reset()
• Open screenshots/live_play/gallery.html in a browser — all 11 cards render, GIFs animate
• Run a TRAIN, then git status — working tree should stay clean

🤖 Generated with Claude Code