One of the choices you've made that I really like is sharing the kernel and userspace filesystem code so directly in the form of libbcachefs. I get the impression this means the kernel can do practically everything userspace can, and vice versa. (I think the only exception is initialising devices by writing a superblock, although the kernel can take over the initialisation of the rest of the filesystem from that point onwards? And maybe turning passphrases into keys for encrypted-fs support which does an scrypt thing?)

As well as giving you really powerful userspace tools for manipulating filesystems, this also suggests that a stripped down busybox module for bcachefs could consist of superblock writing and pretty much nothing else? Maybe a few ioctls to trigger various operations. "Just leave it all to the kernel."