[Zburgers]

Outcome

FlashRL is now a simulator-only, reproducible RL research package with a corrected DQN training contract, exact-budget experiment runner, multi-seed evidence, installable CLI, and live learned-policy demo.

Research result

• 5 independent training seeds, exactly 120,000 frames each
• 100 new held-out evaluation episodes per learned policy
• learned mean: 301.90 (95% training-seed bootstrap interval [270.61, 335.98])
• random baseline: 126.18 (+139.3%)
• rule baseline: 284.17; reported conservatively because it lies inside the learned interval
• independent component study: N-step Dueling Double DQN 289.67, versus Dueling Double 255.79 (+13.2%)

Release evidence

• 90 tests pass from a clean detached worktree
• valid sdist and wheel; installed-wheel CLI/doctor qualification passes
• Gymnasium and Stable-Baselines3 environment checkers pass for state, vision, and hybrid observations
• raw result/config/manifest payload, representative checkpoint, GIF recordings, distributions, and SHA-256 file prepared for v2.0.0
• complete issue-to-evidence matrix: docs/reports/V2_RELEASE_AUDIT.md

Scope decisions

Browser automation and PPO are explicitly deferred and absent from the V2 command/result surface. V2 makes no unsupported browser or PPO claim.

Closes #2
Closes #3
Closes #4
Closes #5
Closes #6
Closes #7
Closes #8
Closes #9
Closes #10
Closes #11
Closes #12
Closes #13
Closes #14