OK the non-ecal-pn tests failed just because I added an extra printout for the BDT score, but all the K-S passes.
Otherwise, for the ecal-pn, there is quite some change:

1. fidiciality changes as expected

• recoil y a little:

• recoil x quite a lot:

Honestly the y distribution with the recoil tracking seems more natural. Sure we expect this to be asymmetric. But I dont think we expect it to be completely on the left side, so having some tracking ending in the right looks OK to me. One thing that's interesting is that the tracking propagation should include both magnetic field and material. I dont actually know how the SP y is made.

2. This will change the projections

• photon separation

• ep angle

looks less peaky in the high end, which to me a good sign

• photon containment energy

• ele cont energy

3. As a consequence the MIP tracking changes too

• lin-reg tracks

I'm kinda surprised this failed the K-S, usually the changes in the tails dont bother the K-S, but ok, from first principles this is expected

• str8 tracks

So here it's a bit weird, I don't know what to make of it. It's not bad for a cutflow tho.

@therwig I thought I'd bring your attention to the validation plots I posted here too. Quite a lot more tracks in the ecal PN as well

I was talking with @danyi211 about this, she pointed out that if the recoil track is on the left, that means that the photon projection is on the right. But if you mismeasure by the tracking that the recoil track is on the right, then the photon projection will be on left -- but that's where the real electron shower is!! so you'll find MIP tracks bc there are MIP tracks from the electron.

Thanks for sharing these - interesting. The straight tracks are quite striking :)
Thinking about your explanation, can we correlate this pathological reconstruction with the electron containment energy? I wonder if some requirement that E/p~1 (something loose, like within ~2 sigma?) or some level of track-to-cluster matching for the electron would clean up this case. We could also look at track quality.
OK, probably you already had some similar ideas about how to fix this, but its gratifying to see a nice explanation here!

We figured out that the pathological events can be killed by requiring 1 recoil track. I have had this comment in the code saying that we need to match the extrapolated track to the CLUE cluster, that will lead to the same results, but that needs the CLUE PR to be merged.

So I propose we merge this in this state and will have the promised improvements as CLUE merges. Please also note the use of the tracking in Ecal is OFF by default. I also plan to include ACTS tracking in the signal + kaon validation configs so we can enable this by default in the future. But my point is that merging this will cause no harm.