Fix scifi states access for BestTrackCreator

The way of accessing scifi states for lepton_id computation was problematic for the long tracks originating from the seeding & matching algorithm, in case the best_track_creator was used. This MR aligns the long track scifi states between Forward and Seeding & matching tracks and ensures that they are accessed correctly when running either individual or combined track sequences.

Closes #428 (closed)

@cmarinbe @dcampora @cofitzpa @tevans @jhorswil

added bug fix rta-hotfix labels

added RTA label

Given that currently the full physics efficiency test is failing because of previous outdated references, I am attaching here the diff that I get locally when running on the nominal Bs2PhiPhi sample with hlt1_pp_forward_then_matching_no_gec_no_ut_validation. - is master and + is with this MR. The main things I notice are:

• The muonID efficiency now goes to around 87% with the forward-withUT and 90% with forward-noUT while also the muon misID + ghost muons are reduced
• Muon line rates are in general lower, which I would think is because of the lower muon ghost rate, given that I'm running on a Bs2PhiPhi sample but better to get the feedback of the BW division study
• Some non lepton line rates are also changed a bit, which I would attribute to lepton_id specific cuts on the filtering of tracks used for the SV reconstruction - perhaps not ideal that SV reco for hadronic lines depends on leptonID performance? (see https://gitlab.cern.ch/lhcb/Allen/-/blob/master/device/vertex_fit/vertex_fitter/src/FilterTracks.cu#L52-55)

diff.txt

/ci-test

added ci-test-triggered label

• [2023-05-19 17:22] Validation started with lhcb-master-mr#7930

Looking at the reference differences, apart from the expected ones thanks to correct muonID, I also see that in some cases the number of reconstructible velo/scifi/long tracks are different, and in particular exactly half wrt master. Running the commands of the test locally, I realized that in master these sequences don't have enough memory and proceed to divide the input slice, which might lead to this duplication of number of tracks and mc particles in the validators? What do you think @dcampora ? In any case, increasing the memory when running from master seems to give consistent results with the fix (which doesn't require any increase in memory), so I would say the fix is also improving on memory usage.

added hlt2-throughput-decreased label

added 1 commit

• b1717751 - Increase GPU memory of CI jobs

Compare with previous version

added 1 commit

• eaea79cf - Update CI references

Compare with previous version

assigned to @raaij

resolved all threads

approved this merge request

merged

mentioned in commit d01e080f

mentioned in merge request !1210 (merged)

mentioned in merge request !1200 (merged)

removed hlt2-throughput-decreased label

mentioned in merge request !1213 (merged)

mentioned in merge request !1216 (merged)