windowFWHM code is failing

Copy of Robert's comment on Slack (auto-guiding channel)

Wilfred Gee and I are seeing a lot of failures in the windowedFWHM code; it seems very fragile.  For example, in pfs_visit_id, agc_exposure_id = 125719, 700288 AG3, the object at xPos, yPos = (668.202640, 136.838380) has flag bit 0x8 set;  running the code from a notebook the problem seems to be from NaN (coming from a negative second moment).  If you take xPos, yPos = (668.01, 137.01) you get a good (and correct) value!

I think that at least part of the problem is that you initialise sx = 1.5;  sy = 1.5 which is too small – it looks as if you took the value from my SDSS code with much larger pixels than the AGs; these values correspond to c. 0.4" seeing.   I think that a value of around 6 is likely to be significantly more stable.  Can you test that using the AG playback on a week or two of data?   You should get many fewer objects with that bit set (of course some of them won't have flags == 0x0 – you'll need to check (flags & 0x8) == 0).   It's possible that this won't be a sufficient fix, or cause something else to fail and we'll have to do some more thinking