The puzzling bit is that replication agents continue happily along, as though nothing were the matter. You have to look in the application event log...
right... so I found a blog that talked about a similar problem and referred to the "useless, truncated message" you get when these plans fail. I liked that 😉