The EDN circuit uses a kind of synchronous filter to get fast settling, despite the relatively low PWM frequency. The synchronous filter can add extra errors from charge injection, which depends on the voltage level or voltage dependent on resistance. So it is kind of elegant, but not so well suited for super higher accuracy.
The more normal way would be a higher order filter - slower, but more predictable. Besides the normal filter function there can be settling due to DA in the capacitors. So settling to the very last ppm will take quite some time. There are chances for a speed-up, but here it starts to get really tricky.
How the main and fine PWM are combined in the EDN circuit is also a little unusual: they add a little to the reference and ground level. The simpler ways and maybe better would be having a second larger resistor for the smaller part, combining at the input of the filter. Having 0.1 % resistors in the combing circuit would limit the extra accuracy too about 8-9 Bits unless an adjustment is done.
Some overlap (thus fine PWM equal to maybe 2-4 LSB steps could be useful for checking the scaling, but more could be more like a problem as it would need more accurate scaling and fine PWM drivers.
So it would be more like 16 Bit coarse PWM, 2 Bits overlap and 16 Bits fine PWM. Thus a theoretical resolution of up to 30 Bits, with some lost in rounding for adjusting the fine scale in software and a lower accuracy than resolution. Even if not absolute linear, the DNL can still be quite good, as PWM is usually (nearly by definition) monotonic.
I would take a look at the circuit for the Fluke 5700 series calibrators. They include a reasonable easy way of compensating for the difference in switch resistance. One could build a similar circuit with modern CMOS switches (e.g. DG4xx) at some places. It still takes a few critical parts: 1 very linear resistor (thus usually low TC and not very small size). A few good caps with low leakage and low DA to keep residual settling times short. A good layout might be needed to avoid RF noise / spikes to cause funny / hard to predict effects. Also parasitic capacitance around the resistor can be a problem. Super fast switching is not the best way as it causes more RF trouble. There is a small range at the bottom and near full scale that is problematic anyway. The good range is more like 1% to 99%.