OpenRead v2.23.0

Added

• A page in a language you have no pack for starts fetching it as it
  loads. The install-time fetch can only guess one pair — English into
  whatever the settings say — because a source language is not knowable before
  there is a page. Everything else was still met the hard way: open a Japanese
  article, press translate, wait out a download that had not started. Now the
  worker asks each page what language it is in as it finishes loading, and
  starts the pack then. By the time a reader has finished the first paragraph
  and reached for the button, it is on its way or already there.

  Fetching all 39 target languages at install was measured first, and rejected.
  Each additional language costs about 130 MB of its own — en→fr +131.7 MB,
  en→de +133.4 MB, en→ar +127.8 MB, on a profile that already held two
  — so the full set is roughly five gigabytes. This extension exists because
  the alternative to it is a server and a five-gigabyte model; shipping one at
  install would be the joke told about us. A page actually opened costs 130 MB
  per language somebody genuinely reads, and nothing for the other 37.

  Asked once per pair per worker, skipped for a page already in the target
  language, and skipped while another pack is downloading — two at once share
  one connection and both finish later than either would alone.

  pnpm e2e:page-pack opens a Japanese article on a profile that has never
  seen one and weighs the profile.