Non-deterministic tracking in pp+pAr reference run test

I see non-determinism in the number of reconstructed tracks when I run Allen on a pp+pAr MB MC sample with pp reference run conditions. See, for example, this failing test:

https://gitlab.cern.ch/lhcb/Allen/-/jobs/42571272

Data file: /scratch/allen_data/mdf_input/ppArBeam2680GeV-expected-2024-MagDown-nu3.5-nu0.37_MinBiasMD.mdf

Geometry: geometry_dddb-20240427_sim10-2024.Q1.2-v1.1-md100

Device: 2080TI

Sequence: hlt1_pp_ref_forward_then_matching_validation

Other Allen args: -n 1000 -m 1100