The funny thing is that these 'garbage characters' are actually numeric characters, but in Cuneiform (as mentioned in T193610). Perhaps the collision is due to improper handling of non-BMP characters in the collation module?
The greatest lie we tell ourselves is 'I'll remember it'.