Fix UB by checking for invalid moves in RL and play_game

[LambertWSJ]

• In play_rl(), skip writing to table when the agent returns -1 to avoid writing out of bounds.
• Do the same in play_game() so negamax/MCTS/RL won't update the board with an invalid move.

[visitorckw]

How did you find this problem?
Was it caught during runtime or purely by code review?

[visitorckw]

IIUC, get_action_exploit() only returns -1 when the board is full, which causes the issue you described. However, in that scenario, play_game() should stop calling play_rl() because check_win() would return 'D'. Therefore, calling play_rl() when there are no empty spaces shouldn't happen.

If this error was observed during runtime, I suspect an implementation error elsewhere is the root cause. Adding a check here would likely just hide the real problem.

[LambertWSJ]

When porting the RL module to kxo[1], I occasionally encountered cases where move = -1. Using dmesg, I noticed that a bit shift had exceeded 31 bits. Because of this, I added last_e to get_action_exploit so that if no best action was found, it would return the last visited position.

Strangely, when I switched back to this commit[1] and removed the max_action == -1 check, the issue disappeared.

[1] kxo - Port reinforcement learning from ttt

[LambertWSJ]

I found that this error occurs when the state space isn’t fully initialized at the beginning during the porting of the RL module to kxo, which can cause move to become -1.

However, ttt initializes the entire state space from the beginning, ensuring that the agent knows every possible board state and preventing the move = -1 issue.

As this issue does not occur in the current implementation, this PR can be closed.