Artifacts when capturing at 96k/sec

[abelian]

Glitches occur as the input signal moves through about +1.3V and -1.3V, as highlighted in the following capture of a 130Hz 5.0 Vpp sine wave:

http://abelian.org/vlf/tmp/180218b.gif

This is channel 1 on the Octo, at 96k/sec with input gain set at the default 38 in alsamixer.

The glitches always occur roughly about +/- 1.3 V on the input, regardless of the digital gain setting. Turning off ADC1/2 HPF makes no difference.

Most noticeable on a low frequency signal. Particularly noticeable when the input amplitude just about peaks above 1.3 or 1.4 V, as in

http://abelian.org/vlf/tmp/180217c.gif

There is no sign of these artifacts when sampling at 48/sec, the signal is clean.
I've tried two specimens of Octo and two Model 3 Pi.
Tried taking the input signal direct to the differential inputs on the connector of the Octo.
All inputs are affected, but the glitches are not in unison across the channels.
Not every transition through +/- 1.3V produces a glitch but most do.
Power supply for the Pi is an officially recommended one.

My application measures parameters of a pulse and these glitches are significant,
here's an especially bad one.

http://abelian.org/vlf/tmp/180217b.gif

Hoping there's a solution for this.