Here's an amusing example, where we find the muon between the traces.

 is an example of arc lines being labelled as CRs (this is using makeCRMosaic from PIPE2D-773.  Each CR appears in two panels, raw on the left, calexp on the right.  The axis labels are in mosaic units, but the utility provides a mouse-over in calexp coordinates).  This is with repair.interp.modelPsf.defaultFwhm=2.5.  If I lower this to 2.25 the problem goes away, but we seem to miss real Cosmics

Here's a rather worrying presumed-cosmic ray .  It's flagged as CR (green pixels at the bottom) (but that may just because it's so bright, and we "should" be labelling the sky lines too).  The cross-section is slightly wider than an arc line ( ), so it's going to be hard to get things like this out of the data (although it's centroid isn't consistent with being on any fibre)

I think combining multiple exposures is going to be an important part of our CR mitigation strategy, because we're not going to be able to find all of them based on morphology alone.

I don't disagree with Paul, but we may not always have the luxury of CR splits.

Here's an example of CRs we're missing with repair.interp.modelPsf.defaultFwhm=2.25

And here's another fat trail (note the muon tracks in the background). I'm beginning to worry if we're not fully depleting the CCD, as I don't remember things like this from HSC (or LSST, but it has thinner devices and smaller pixels).

Here's another fascinating event:

Here's a case of a CR masquerading as part of an emission line.
Here we detected the CR, but there's flux in the extracted spectrum; here's the spectrum, but only the blue-side is labelled as a CR.

Here are three of these fat events in one frame 

And here's a CR that's just to the left of the trace at the position of an emission line. 

Do you have an algorithm in mind for finding these CRs, or is this a research project?