So I can reproduce 7 working and 8 not working in my hands with your first training model, at least not working over a scale of ~3-4 hours. If I use one of our old training models (specifically the one from this example workflow), both images run but each real worm is broken into many tiny "worms". So at least in part the issue is with your training model.
I think the overarching issue however is that your images are just really big and really high-magnification compared to what these modules were designed for, which eats a lot of memory; consequently, the worm models you've trained are also huge, and the combination of trying to fit huge worms to huge images is straining the limits of the system. When I downscaled your images to 0.1 size with the "Resize" module then tried to run with the CP worm model, both images ran pretty quickly and were fit correctly to my eye. I don't know enough about your upstream or downstream workflows to know if your images "need" to be that big to capture something else that you care about (like your fat droplets) but I suspect you may be able to find a sweet spot where you can downscale the size enough to get images that are small enough to train and to be processed efficiently but still large enough to find the features you care about.
I hope that helps!