fix: correct ground_truth comparison in BanditEnv

[bigsawman]

Summary

• ground_truth in BanditEnv is a dict mapping button names to probabilities (e.g. {"red": 0.6, "blue": 0.3, ...}), but the final-turn winner check compares a button string directly against this dict (button == self.state.game_state['ground_truth']), which always evaluates to False.
• This means the player can never be recognized as having chosen the correct button — every game ends with an incorrect outcome and a regret-based reward.
• Fix: find the button with the highest probability via max(..., key=...), then compare the player's choice against that.

Reproduction

ground_truth = {"red": 0.6, "blue": 0.3}
button = "red"
print(button == ground_truth)  # False — string vs dict, always False

Test plan

• Verified that _regret() already correctly uses ground_truth as a dict (calls .values() and indexes by button name), confirming this is a dict, not a string.
• Confirmed the fix matches the intended semantics: reward 1.0 when the player picks the highest-probability button.

[borgr]

Ready to merge. Correctly diagnosed: ground_truth is a dict {button: prob}, so the old final-turn check button == self.state.game_state['ground_truth'] (str == dict) was always False — the reward=1.0 branch was dead code, and even a correct pick fell through to the regret branch. The fix max(ground_truth, key=lambda b: ground_truth[b]) is the right one and matches how reset() builds the distribution (the intended button always gets the highest probability, so the argmax is exact). Minimal and in-scope.

[borgr]

Verified. ground_truth is a dict {button: prob}, so the old button == self.state.game_state['ground_truth'] compared a str to a dict — always False. The +1 "correct button" reward was therefore unreachable, and even the optimal button fell through to the regret branch (scoring 0.0, its own regret). Taking argmax over the probability dict is correct: by construction (reset assigns the ground-truth button 0.5 + p_gap/2 and every other button < 0.5 - p_gap/2) the ground-truth button is always the unique highest-mean button.

LGTM, recommend merge.