g2048 environment: new tile logic is bugged (different from the original)

[bbbbbbbbba]

In the c_step function, the function place_tile_at_random_cell is called after move, but before update_stats_and_get_heuristic_rewards. This is a problem because place_tile_at_random_cell relies on game->empty_count, which is not updated after a move merges some tiles, only in update_stats_and_get_heuristic_rewards.

PufferLib/pufferlib/ocean/g2048/g2048.h

Lines 558 to 572 in 5cbaf88

	void c_step(Game* game) {
	float reward = 0.0f;
	float score_add = 0.0f;
	unsigned char prev_max_tile = game->max_tile;
	bool did_move = move(game, game->actions[0] + 1, &reward, &score_add);
	game->tick++;
	if (did_move) {
	game->moves_made++;
	place_tile_at_random_cell(game, get_new_tile());
	game->score += score_add;
	// Add heuristic rewards/penalties and update grid stats
	reward += update_stats_and_get_heuristic_rewards(game);
	reward *= game->reward_scaler;

PufferLib/pufferlib/ocean/g2048/g2048.h

Lines 204 to 221 in 5cbaf88

	static inline void place_tile_at_random_cell(Game* game, unsigned char tile) {
	if (game->empty_count == 0) return;
	int target = rand() % game->empty_count;
	int pos = 0;
	for (int i = 0; i < SIZE; i++) {
	for (int j = 0; j < SIZE; j++) {
	if (game->grid[i][j] == EMPTY) {
	if (pos == target) {
	game->grid[i][j] = tile;
	game->empty_count--;
	return;
	}
	pos++;
	}
	}
	}
	}

In update_stats_and_get_heuristic_rewards:

PufferLib/pufferlib/ocean/g2048/g2048.h

Line 481 in 5cbaf88

game->empty_count = empty_count;

The consequences are:

1. When the grid was full before a valid move, the move will necessarily have merged some tiles, but place_tile_at_random_cell still sees that game->empty_count is 0 and returns immediately. Therefore the game does not spawn a tile for that move.
2. When the grid was not full before a move merged more tiles, place_tile_at_random_cell will see a smaller value of game->empty_count than the actual empty count, and thus will never spawn a tile in the last empty cell(s).