Polyseat 0.24.0

A seat never took anything from the machine it runs on but its hardware.
Not its language and not its keyboard layout, so a German host handed out
seats that answered in English and put a US layout under every key. Both come
from the host now, and a bug found on the way out stops the shared library
carrying a folder back and forth forever.

Seats have to be built again. The recipe is at generation 43, so every
seat is marked stale and wants provisioning. A few minutes per seat.

• The seat takes the host's language. A seat is built from a plain Arch
  image, and a plain Arch image has no locale: systemd falls back to
  C.UTF-8. Steam and the launcher come up in English, and so does every
  Windows title in the seat, because wine reads the Unix locale to decide what
  GetUserDefaultUILanguage answers and a game that ships fifteen
  translations picks by that alone. The new locale step reads LANG out of
  the host's /etc/locale.conf, generates that locale in the seat and writes
  the seat's own locale.conf.

  Read from the file rather than from the daemon's environment, because
  polyseatd is a system service: its LANG is whatever systemd started it
  with, not what the person sitting at this machine chose. C, POSIX and
  C.* are the absence of a locale rather than a choice and are not copied.

  Two things are worth knowing about when it takes effect. A running session
  keeps the environment it started with, so the language arrives when the
  session next starts. And wine writes a prefix's idea of the locale when the
  prefix is updated, not on every run, so a prefix built before the seat had a
  language keeps answering with the old one until wineboot -u is run in it.

• The seat takes the host's keyboard. xkb's default is us, so signing in
  to a store from a seat meant hunting for y and z. The layout, variant,
  model and options come out of the host's /etc/vconsole.conf, which is
  where systemd-localed keeps them, and become an input * block in the
  seat's sway config. KEYMAP is deliberately ignored: that one is the
  console keymap and a container has no console to apply it to.

  A host that never set a layout gets no block at all rather than an explicit
  us, so sway keeps its own default instead of being handed a guess. A
  layout is also required before a variant, a model or an option is carried:
  half a keyboard would replace the default with something incomplete.

  These values are written into a config file that is read line by line, so
  they are checked before they go in. A value carrying a newline or a brace
  would not be a bad layout, it would be a sway directive somebody else wrote.

• A cloned symlink gets its own timestamp back. Clone restored the
  modification time of every file and every directory it copied, and of no
  symlink. A folder's version in the pool is the newest timestamp anywhere
  inside it and measure stats symlinks along with the rest, so a copy
  containing a single symlink always measured newer than the original it was
  made from. The pool then handed the folder back and forth for as long as it
  ran: the seat that had just been written from the pool immediately looked
  like the one holding the newer copy, was taken back into the pool, and was
  written out over the other seat again a minute later.

  This is the failure the times are preserved to prevent, described in
  docs/architecture.md as two copies carrying each
  other back and forth forever. It was closed for files and directories and
  open for symlinks.

  It surfaced putting a game into shared/ whose writable state has to stay
  per seat, which is a symlink out of the tree and therefore the first folder
  in this library to contain one. What it looked like was not an error
  anywhere: files written into the shared folder were simply gone a few
  minutes later, in every seat at once.

  os.Chtimes was not an option and is the reason this was missed. It follows
  the link, so it would have stamped the target, and for a link pointing out
  of the pool the target is not the pool's to touch. utimensat with
  AT_SYMLINK_NOFOLLOW is the one that does not follow.