It resulted in random, hard to reproduce ENOSPC errors down the line without either data or metadata being anywhere close to full. Neither us nor the btrfs developers that took a look at it were able to figure out what exactly went wrong, but it was something about new blocks not fitting anywhere despite lots of free space.

Someone on #btrfs said that the filesystem layout is a lot different when using the conversion tool and all of the regression testing happens with regular filesystems, not converted ones.

We reinstalled all machines from scratch. Never happened again.