Does every row in the CSV have a unique ID of some sort? Export what you have in AGOL and isolate the missing records and see if there's some notable difference between them and the successful records. You might have to open the file in a text editor to see exactly how the data is stored.
I would hazard a guess that there is something not right with the coordinates for those missing 51 records. One approach you could use is to join the spatial layer back to your original .csv - rows with nulls showing on the spatial layer side of the join will be your missing .csv values and you can look at the coordinates in that context.
