Your filter input impedance is the problem. That is much lower than the 100 ohm R load, such that peak excursions in the PWM are clipped and the midpoints have shifted as a result.
You need to define both the s11 and s21 response specs or the input Return Loss and transmission loss from loading and also the output impedance of your transformed RdsOn Bridges. Then and only then can you design a filter to meet your specs.
Consider the impedances of source and load at both line and switching frequency in this operation. Then define your impedance requirements so that the no load filter current is minimal <<10% rated current. A different topology is likely needed that raises series impedance above 10kHz while not affecting source impedance at line f.