The environment is the game itself, the fixed rules of Hokm that every move happens inside. Four players in two partnerships, a 52-card deck, a trump suit chosen at the start of each hand, and the one rule that gives the game its shape: you must follow the led suit if you hold it, and may only trump or discard when you cannot. The engine deals, tracks whose turn it is, decides who wins each trick, and scores the hand.
These rules are a frozen node. They sit outside the learner's reach on purpose, because they are exactly what an optimiser would be tempted to bend if it could. An agent that could quietly redefine a legal move would soon be winning a game no one else is playing. So the rules are enforced by the environment, not learned by the policy: the agent may only choose among the moves the rules already allow. That separation is what makes a trained result mean anything.