Is this happening on a Linux system?
Does the fiducial show up in the MRHead eye at the end of running Testing -> Test Cases -> FiducialLayoutSwitchBug1914?

I used Windows. The test you mentioned shows the fiducial but fails

Running the aglorithm
Conventional view
Getting MR Head Volume
<b>File already exists in cache - reusing it.</b>
<b>Requesting load</b> <i>MRHead</i> from C:/Users/pinter/AppData/Local/Temp/RemoteIO/MR-head.nrrd...

<b>Load finished</b>

Placed a fiducial
Panned and zoomed
Starting seed display coords = 317, 21, 0
Fiducial starting position
Red Slice only
Conventional layout
Ending seed display coords = 269, 19, 0
Fiducial ending position
Difference between starting and ending seed display coordinates = 48.0416
Display coordinate difference is too large!
Expected < 1.0 but got 48.0416
Test failed!

Hrm, the test also fails on my Mac, but with a much smaller difference (16), I have a feeling that it should be scaled to the Slicer window size.
I was finally able to reproduce the panning disappearing fids, but only if I panned so that the fiducial was very very far off screen, and only in certain directions in certain windows (about a full window's width off). There might be something in the light box code that's getting triggered since the RAS to display calculation could be thinking that it's so far off that it should be in the next lightbox view, but I have to dig into it and see why it's not coming back when you pan back - looking at the fix for bug 1914, there could be a CTK issue.

Interesting. For me, even if I move the fiducial barely outside, it doesn't appear when I move it back.

Out of curiosity, what's the resolution of your monitor? And how big do you have Slicer set to be? These are the results of the python commands I have with my monitor set to 1440x900:

slicer.util.mainWindow().size
QSize (1051, 768, at: 0x7fa6cfa1be80)
lm = slicer.app.layoutManager()
lm.sliceWidget('Red').size
QSize (174, 329, at: 0x7fa6cfeb8850)

1920*1080, Slicer is full screen (account for header and taskbar)

But it doesn't make a difference for me. As I said even a few pixels out of view, lost forever (until hidden/shown again)

I was able to reproduce it on Mac relatively easily, although it is not as deterministic as on Windows.
If you move the markup in and out a few times (max 8), it disappears. You don't even need to drag it far away.

This doesn't seem to be reproducible anymore on Mac or Windows.

Thanks!
Confirmed, it works well now for me too.
I'm closing it as fixed, as a commit definitely fixed it at some point. Not sure which, but it doesn't matter much now.

I also confirm the issue is not reproducible on Ubuntu 14.04 (debug build based of r24129)

Jc, did you reopen on purpose?

@Csaba: I guess I also marked the issue as resolved. Closing it again now... sorry for the noise.