Improve AI difficulty scaling.

internal/ai/ai_test.go

@@ -254,6 +254,26 @@ func TestSimulateBattleIsPure(t *testing.T) {
 	}
 }
 
+// TestBlunderCurve pins the difficulty calibration's shape: only sub-competent
+// bots ever throw a game on purpose, and they do so more the weaker they are.
+// Medium (== competentLevel) and hard must never blunder, or they would not be
+// the even-match / favourite the difficulty design promises.
+func TestBlunderCurve(t *testing.T) {
+	if p := blunderProb(competentLevel); p != 0 {
+		t.Errorf("competent bot blunders with p=%.3f, want 0", p)
+	}
+	if p := blunderProb(1.0); p != 0 {
+		t.Errorf("hard bot blunders with p=%.3f, want 0", p)
+	}
+	weakest, easy := blunderProb(0), blunderProb(0.25)
+	if !(weakest > easy && easy > 0) {
+		t.Errorf("blunder rate must rise as level falls: level0=%.3f easy=%.3f", weakest, easy)
+	}
+	if weakest > 1 {
+		t.Errorf("blunder probability %.3f exceeds 1", weakest)
+	}
+}
+
 // TestDecideShopNeverSellsLastPet guards the invariant that the bot never
 // voluntarily turns its whole deck into food. In a hopeless late-game spot
 // (final round, a strong modeled opponent, so every rollout is a loss) the