hosts 1.0.1

Verification of what 1.0.0 already claimed, and one claim that turned out not
to be true.

Fixed

• Concurrent writers no longer lose each other's work. The lock was taken
  inside the write, after the file had already been read, so two writers each
  computed their change from the same starting point and the second one
  discarded the first. Serialising the writes prevented nothing, the rename
  having already made those indivisible: the update that got lost was lost
  during the read. The lock is now taken before the file is read and held
  until the new content is in place. 0.2.0 documented the protection this
  provides; it did not provide it.
• hosts edit holds that lock for as long as the editor is open, so another
  writer waits and then fails saying so rather than quietly discarding one of
  the two changes.
• The apt index is rebuilt after a release again. It was chained to the
  release event, which a release created by a workflow using the built-in
  token never raises: GitHub suppresses those so that workflows cannot
  trigger each other in a loop. It now chains to the release workflow itself,
  which is the mechanism meant for it. The index for 1.0.0 was published by
  running the workflow by hand.

Added

• Tests that kill a write partway through, at a different point on each pass,
  and require the file to hold either the whole of the old content or the
  whole of the new one. That guarantee is the one the whole program rests on
  and had only ever been reasoned about.
• A test that runs two writers at once against a file large enough for the
  race to be certain rather than merely possible. It is what found the lock
  defect above.
• Tests that put the parser against input it was not written for: control
  bytes, addresses and names that nearly parse, lines of only whitespace, and
  a very long line. Three invariants are checked, the strongest being that
  export gives back exactly the bytes that were read. The generator is
  seeded so a failure can be reproduced. Nothing was found, which is also
  information.
• A portability job that runs the whole suite on Debian 11 and 12 and Ubuntu
  22.04 and 24.04, as an ordinary user rather than as root so that the tests
  about refused permissions do not skip themselves. Everything had until now
  only ever run on one Ubuntu, which left every assumption about stat,
  date, realpath, sync, flock and install untested.

Changed

• The README no longer implies bash 4.4 is tested. It is the documented floor;
  what the suite runs against is bash 5.1 and 5.2, because nothing still
  supported ships 4.4.
• A test no longer expects a file to come out mode 644. Debian defaults to a
  umask of 022 and Ubuntu to 002, so the same fixture is 644 on one and 664 on
  the other; the sidecar records the mode the file actually has, which is what
  the test now checks. The portability job found this on its first run, which
  is the sort of thing it was added for.