BTW I think your DT-880 measurements are messed up too.

Well, that is not impossible, maybe the microphone "corrects" the distortion by having second order distortion with inverse phase relative to the headphone drivers. At high SPL and high frequency, the measurement definitely does become microphone limited (if I move the microphone away from the drivers, so that it does not receive as high SPL, the distortion drops significantly). However, the result seems to be fairly consistent with changing frequency, level (graphs will be shown soon), and even with different headphones I still get low distortion.

One thing I cannot guarantee is that the SPL is correct. I simply used Tyll's sensitivity measurements as a reference, so that I only needed to measure voltage, which I can do accurately enough.

You are getting way too low second harmonic for the DT880 driver or any dynamic driver. In my years of experience, I don't recall seeing any driver have a second harmonic at -68db at any reasonable volume.

It is actually fairly consistent with the HeadRoom distortion measurements, at least as far as the 2nd harmonic is concerned. For some Sennheiser headphones, HeadRoom in fact shows even lower distortion.

I know Tyll's IF measurement ~0.3% @1kHz is THD+noise, but THD+noise is typically only a little be higher than THD by itself, not a magnitude higher.

0.3% THD+N at 90 dB SPL is 39.5 dB distortion + noise SPL. Without A-weighting (and THD+N measurements are usually unweighted), it can plausibly be entirely noise. Especially if the possibility of inaccurate SPL (due to the headphone not having perfectly flat response) is taken into account, so it may even be a few dB less than 90.

As an example, the THD+N (unweighted 20 Hz to 20 kHz) of the same tone that I plotted earlier to show the -68 dB distortion is about 1%. Of course, that was recorded with a crappy Behringer measurement microphone in a "normal" room, but even with better equipment it is not easy to get the noise at 90 dB SPL to under 0.1% without weighting.

The fact that the InnerFidelity graphs so often show less THD+N at 100 dB SPL than at 90 dB also makes it likely that the lower SPL measurement is actually dominated by noise.

Subtract the noise, and the THD is likely to be ~0.15 to 0.25%, with the second order distortion being similar to THD (since second order dominates), not your 0.04%.

Of course, if there is actual high THD, then adding the noise will not increase it as much. But the point is that even the noise alone can already produce a relatively high THD+N, so it masks any really low THD. Even for your example of 0.25->0.3% increase, 0.166% noise is needed.