I would bet that with a 0V difference on A+/A-(plus ADC noise and other noise), due to the floating ground, the HX711 would produce raw values like random(-10,10) insensitive to any weight on the scale, and calibrating between a set of samples at taken when the scale is loaded, and another set when unloaded would just give a random slope. The calibration routine would still run and report itself as successful, but the scale determined by the routine would be random.