After some further searching I found out how these 幽霊文字 (the "ghost characters") came about. When the JIS set was being decided on, the creators of the set had to work within certain limits, so decisions about which characters to put in the set and which to exclude were of crucial importance. The compilers especially wanted to make sure that as many Japanese place names as possible (many of which use rare or archaic kanji) were representable on the computer systems. To do this they consulted a number of different records from various places, but in doing this, some characters were transmitted incorrectly, thus creating "new" characters.

Apparently the records from these compilers were not well preserved and so it's not really known absolutely for sure which characters are ghost characters and which are not. A small number of people have done research on the issue and been able to find actual instances of use of certain characters that were thought to be ghost characters.

For others, the source of the error has been discovered. For instance, the non-character 妛 seems to have come about from a sheet of paper with 山女 written on it, vertically, with what looked like a line representing a stroke between them. It turned out later that the "line" was actually a crease in the paper that, when photocopied, made a shadow that looked like a stroke. Another character was tracked down to someone misreading a scribbled character that another person had written on a sheet of paper.

This reminds me a little bit of the HTTP protocol, which contains a field called a "referer" -- a misspelling of "referrer" that somehow made it into the standard definition, and which now must be used by all programs utilizing HTTP. I actually encountered a bug with this in a summer program I was working on. We were having all kinds of problems implementing a web proxy, and it turned out that one of our major bugs was that we were using "referrer" instead of "referer" in the program.