research: freebsd-livecd-gunion-reroot — experimental architecture plan

Plans a clean-sheet experimental livecd architecture that combines
gunion (block-level RAM overlay) with reboot -r reroot, sidestepping
the chroot pattern that breaks kernel-context firmware loading on
all our existing livecd variants.

Marked EXPERIMENTAL — not yet built. Intended as a separate sibling
repo (freebsd-livecd-gunion-reroot) so we can iterate without
disturbing freebsd-livecd-gunion / freebsd-livecd-unionfs /
freebsd-launchd, all of which work today using a chroot pattern.

Key technical content:

- The kernel-vs-userspace namespace split that breaks firmware
  loading on all three current livecds (verified empirically on
  Lenovo with iwlwifi-8000C-36.ucode "File size way too small!").

- The naive gunion+reroot pattern fails because mdconfig holds the
  rootfs.uzip vnode busy on cd9660; vfs_unmountall() can't unmount,
  vfs_mountroot() lands in inconsistent state, panic risk.

- Two ways to dodge: (A) loader-preload rootfs.uzip into RAM-backed
  memdisk via loader.conf mfs_load="YES" — mfsBSD canonical pattern;
  or (B) in-init.sh malloc-md copy + vnode-md detach. A is preferred,
  B is fallback if loader has size limits.

- Why gunion specifically: it's the only writable-everywhere overlay
  that can be a reroot target. unionfs can't (per unionfs(8) "it is
  meaningless to mount unionfs over a root mount point"). tmpfs root
  burns RAM. RO UFS + tmpfs subdirs is the mfsBSD pattern but loses
  the "everything writable" semantic.

- 6 phases: scaffold, loader preload, gunion+reroot in init.sh,
  firmware-load verification on real hw, CI boot-test extension,
  comparison vs the symlink trick in freebsd-launchd commit 51e1b80.

- 5 documented risks: boot-path regression, loader size limits,
  gunion perf, RAM floor (600-800MB), half-rerooted state.

- 6 open questions, including: does loader preload 400MB cleanly?
  does shell-as-PID-1 calling reboot -r work? symlink trick vs
  reroot — which wins on the head-to-head?

Linked into FreeBSD Research as a sibling of freebsd-livecd-plan
(not nested under launchd subtree — this is a livecd architecture
exploration, parallel to existing livecd plans).

Whether this gets adopted depends on Phase 5 comparison with the
symlink trick. If symlink solves the firmware problem with much
less effort, this stays as a reference document. If reroot has
meaningful additional benefits, freebsd-launchd may eventually
migrate to the same model.

Author: pkgdemon