Test that do not use l2-input (they use ip4 or ip6 instead) are not affected.
As device tests are functional, single loss means failed test. This is why 1n-spr device jobs are currently non-voting.
The packet sometimes appear in packet trace, sometimes not, so it can be lost at multiple places, but the correlation with L2 suggests the loss happens on VPP side.
Allegedly, this is a driver issue, and was affecting also earlier Xeon testbeds, but not 1n-skx for some unknown reason.