The guy uses 0.01% accuracy resistors (you can also order one made with 0.02% and 0.05% accuracy resistors)
AFAIK (but may be wrong here)
This may yield an ENOB of around 13 bits
The 0.05% ladder may even end up with just 11 bits ENOB (depending on the actual accuracy of the MSB and 2SB resistor)
To create a ladder that has 16 bit accuracy the MSB resistor has to have 0.0015% accuracy
To create a ladder that has 24 bits accuracy the MSB must be  0.00006% accurate
The culprit lies in the 0V audio line where 100000000000000000000 switches over to 011111111111111111111111  (this is for 24 bits)
All those 011111111111111111111111 bit values added must sum to the same value as a single MSB (-LSB)

Now you know why it is so hard to create a good (high ENOB/accuracy) ladder DAC chip and why the ones with a high ENOB are expensive.

undoubtably the DAC will sound good though.

He says this in the first scentence:

It's a DAC module based on a discrete R-2R sign magnitude DAC design

When he uses 0.01% in the MSB to 3SB positions and uses 0.05% for the rest he may well get decent results.
I did kind of 'miss' the sign magnitude part though.
The sign magnitude part is where the DAC never crosses the 100000000000000000000 to 011111111111111111111111 border so no crossover type of distortion but there still will be the linearity problem.
It is basically 2 R2R ladder DAC's departing from the LSB, one in the negative direction and one in the positive direction of which the signals are combined.
This can also be done with 2 of the shelf ladder DAC chips per channel, but takes a conversion stage so the DAC sections receive the correct bit pattern.
The non-linearities are simply found in other parts of the INL line, just not in the middle any more also the tolerance issues of the MSB are still equally relevant.

The pricepoint is fun though.

I have other worries as well 

Does INL/DNL really matter for audio? It is a static measurement. Don't know if those plots make sense for S-D DAC anyway.

This is what ultra and I have mentioned before. The reason you don't see these plots on S-D DACs is because it's not considered a relavent spec for their applications. For other applications, it can be critical.

Although it would be interesting if pirates found a correlation between that spec and perceived sound quality.

It can take a long time depending on how one approaches the problem. I can also see INL/DNL measurements being a bit off for delta sigma. Unlike with ladder DACs, if one puts different waveforms to test linearity one might get different results.

Say we use a ramp going from low code to high code, and then another ramp going from high code to low code. Even if we held the ramp values for a while to account for the fact that delta sigma is bandlimited after the LPF filter (otherwise it's just random 1's and 0's/-1's) and therefore not "static", either ramp waveform might yield different linearity results. This perhaps because of the behavior of the DS accumulators (memory). Most designers obviously just go for other metrics.

There might be a way to properly do this, but I don't know how yet. Maybe ignoring the possible waveform dependence and just using a ramp from low code to high code and waiting a few ms per capture is sufficient. There are some cases where delta sigma INL is reported, but mostly in ADCs (not DACs). From the TI military applications catalog: http://www.ti.com/lit/sg/sgzt005/sgzt005.pdf

Consider the ADS1258-EP 24-bit 125 kSPS 24-bit delta sigma:
http://www.ti.com/lit/ds/symlink/ads1258-ep.pdf
Reported INL is 0.0003%, or 2^24*0.0003%/100 = 5 LSB INL

Consider the ADS1278-EP 144kSPS 24-bit delta sigma (about 18 bits of resolution)
http://www.ti.com/lit/ds/symlink/ads1278-ep.pdf
Reported Typical INL is 0.0003% (5 LSB)

Not all military app delta sigma ADC report INL. Don't think PCM4202-EP does.

There is a military audio DACs in that document: PCM4104-EP. It is a delta sigma 24 bits, and it does not report INL information. It is about 19 bits of resolution.

These are proly mostly for audio applications, and not necessarily high rate which might require a different architecture. I can see however a different architecture successfully used in audio applications if sufficient resolution is provided in the audio band.