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

.gitignore

@@ -7,15 +7,27 @@ __pycache__/
 .Python
 
 .vscode
+.cursorindexingignore
+
+# Claude Code worktree config — per-session local state, not part of the project
+.claude/
 
 # Jupyter Notebook
 .ipynb_checkpoints
 
 # Sandbox additions (not upstream)
-.venv/
+# .venv as a bare name covers both a real directory and a symlink to a
+# shared parent venv (a trailing-slash pattern misses the symlink case).
+.venv
 recordings/
 # tensorboard event files live in subdirs of runs/ (e.g. runs/<exp>/events.out.tfevents.*)
 runs/**/events.out.tfevents.*
+# per-training-job stdout logs written by frontend/server.py on /api/training/start
+runs/_training_*.log
+# Per-run summary CSVs that deep_rl_zoo.<algo>.run_atari appends to on
+# every training run. Untracked so any TRAIN keeps the working tree
+# clean — historical snapshots remain in git history if needed.
+logs/*_atari_results.csv
 # new training checkpoints; bundled ones are kept via !-rules below
 checkpoints/*.ckpt
 !checkpoints/IQN_Pong_2.ckpt

deep_rl_zoo/gym_env.py

@@ -30,6 +30,15 @@
 import os
 import datetime
 import numpy as np
+
+# NumPy 2.0 removed np.bool8; gym 0.25.2 (pinned by this repo) still references
+# it in its passive_env_checker.py during env.step. The upstream README §9
+# notes this as a known deprecation warning, but on NumPy 2.0+ the attribute
+# is gone entirely and gym blows up. Restore the alias so every entry point
+# (run_atari / run_classic / eval_agent / frontend.stream_eval) is covered.
+if not hasattr(np, "bool8"):
+    np.bool8 = np.bool_  # type: ignore[attr-defined]
+
 import cv2
 import logging
 import gym

frontend/stream_eval.py

@@ -39,6 +39,12 @@
 from typing import NamedTuple
 
 import numpy as np
+
+# NumPy 2.0 removed np.bool8; gym 0.25.2 (pinned by this repo) still references
+# it in its step/reset paths. Restore the alias before any gym import.
+if not hasattr(np, "bool8"):
+    np.bool8 = np.bool_  # type: ignore[attr-defined]
+
 import torch
 
 import gym
@@ -99,7 +105,7 @@ def _build_iqn(state_dim, action_dim, device, checkpoint_path, env_name) -> tupl
     return network, actor
 
 
-def _build_prioritized_dqn(state_dim, action_dim, device, checkpoint_path, env_name) -> tuple:
+def _build_dqn(state_dim, action_dim, device, checkpoint_path, env_name) -> tuple:
     network = DqnConvNet(state_dim=state_dim, action_dim=action_dim)
     ckpt = PyTorchCheckpoint(environment_name=env_name, agent_name="DQN", restore_only=True)
     ckpt.register_pair(("network", network))
@@ -114,6 +120,25 @@ def _build_prioritized_dqn(state_dim, action_dim, device, checkpoint_path, env_n
     return network, actor
 
 
+def _build_prioritized_dqn(state_dim, action_dim, device, checkpoint_path, env_name) -> tuple:
+    network = DqnConvNet(state_dim=state_dim, action_dim=action_dim)
+    # The bundled PER-DQN_Pong_4.ckpt is stamped with agent_name="PER-DQN"
+    # (matches deep_rl_zoo/prioritized_dqn/run_atari.py line 96). The previous
+    # "DQN" value here caused PyTorchCheckpoint.restore to reject the file with
+    # 'agent_name "PER-DQN" and "DQN" mismatch.'
+    ckpt = PyTorchCheckpoint(environment_name=env_name, agent_name="PER-DQN", restore_only=True)
+    ckpt.register_pair(("network", network))
+    ckpt.restore(checkpoint_path)
+    network.eval()
+    actor = greedy_actors.EpsilonGreedyActor(
+        network=network,
+        exploration_epsilon=0.01,
+        random_state=np.random.RandomState(0),
+        device=device,
+    )
+    return network, actor
+
+
 def _build_rainbow(state_dim, action_dim, device, checkpoint_path, env_name) -> tuple:
     atoms = torch.linspace(-10.0, 10.0, 51)
     network = RainbowDqnConvNet(state_dim=state_dim, action_dim=action_dim, atoms=atoms)
@@ -145,6 +170,7 @@ def _build_ppo_rnd(state_dim, action_dim, device, checkpoint_path, env_name) ->
 
 
 ALGO_FACTORIES: dict[str, Callable] = {
+    "dqn": _build_dqn,
     "iqn": _build_iqn,
     "prioritized_dqn": _build_prioritized_dqn,
     "rainbow": _build_rainbow,