Some late observations.
Ages ago I remember reading a commentary from Kevin Gilmore about his design philosophy. He seems to be one of the low global feedback adherents. Looking at a photo of the PCB, one can get a pretty good feel for the design, and this is confirmed by some of the description made about the circuit on the HeadFi web site. It is a dual differential pair design, using JFETs in the differential pairs. One can see the pairs glued back to back, the current sources (which appear to be biased by LEDs). There is some confusing commentary about a servo current source, which would be the op-amp present on the PCB. I can't see how this would be running the current source, but it could well be a DC offset servo. Then four voltage amplifier transistors in a symmetric setup (probably emitter followers) and then the four pairs of paralleled output devices. So a conventional push-pull class-A output stage. This is a high performing setup, but perhaps the choice of lower feedback is the reason for medium THD. The web site's claimed performance almost exactly matches Amir's.
But, more pertinent, a lower feedback class-A design is not very robust to power supply issues. A less noisy PS may well improve the result. The new PS is $300. Given the Topping A90 already wipes the floor with this amp for $500, I'm not convinced spending more money on the PS is good value, and $800 for the amp/PS combination looks very poor value as a new purchase.
A hidden question with any class-A output stage is the bias current. No class-A output can stay class-A forever. One the delivered current exceeds the bias current operation ceases to be class-A. A single ended class-A will just clip. A push-pull class-A will revert to class-AB. This is why impedance of the load is important. Any class-A amp really should include a minimum load impedance in the specs to indicate the load when it no longer operates in class-A at full swing. Luckily many low impedance headphones are also high sensitivity, so they don't need large voltage swings, and may remain class-A in normal use. But a low impedance, low sensitivity headphone may well find itself taking the amplifier out of class-A operation.
Most amplifiers will simply linearise the output with global feedback as they transition through the A to AB crossover point. A low global feedback design will tend to show more distortion as the amplifier does this. One suspects that is part of what we see here in the measurements of distortion versus voltage swing and load. There are a number of break points on each curve, and these likely represent different transitions of the output stage operation as it swings.
This isn't to say that class-A is deficient, indeed it is ubiquitous in such circuits, but one needs to be clear about its limitations. In the best of worlds it always provides superior performance, but the limitations of its application must be taken into consideration. (I note that the much higher priced Gilmore amps use what appears to be the same building block, but with added heat sinking, so I wonder if they push the bias current up.)