What strategy breaks the game?

Nemotron Nano proposed 32 strategies over 2 generations; the engine played 30 rounds of each against the CPU (960 rounds). Measured: the top strategy's landed hits were SWEEP 117, GROUND 49, BODY_JAB 16, PUNCH 7, KNEE 6. All three strategies exploit the CPU’s tendency to retreat after being hit by sweeps—specifically, the Sweep move triggers the CPU’s decision logic where `if (spammedSweep && dist <= 1.25 && grabReady)` leads to a THROW_PERSON action, but more critically, the repeated use of sweep causes the CPU to enter recovery or block states where it fails to defend against follow-up grabs due to its state-based retaliation rules not accounting for rapid re-engagement after sweep-induced stun or knockdown. The actual damaging move in each case is the ground-based grab or throw (implied by the GROUND hit count), which lands because the CPU, after being swept, enters a RECOVERY or BLOCK state where its decision function prioritizes retreating or parrying over blocking low or defending against throws, especially when `player.state === 'RECOVERY'` leads to a CROSS/SWEEP/ROUNDHOUSE response that misses if the bot is already in grapple range. The guard rule further confirms this: `isGuarding` only blocks low sweeps if the defender is crouching, but the CPU rarely crouches after being hit by a sweep, leaving it vulnerable to standing throws or ground attacks that bypass its high-block focus. Thus, the bot’s sweep spam manipulates the CPU’s memory and state tracking to create openings for unblockable grabs.