The page path has reported that download since 2.15.0, whose changelog
entry reads "two minutes of silence on the first use of a language pair
is the bug this replaces". The PDF path, added three versions later,
never passed the callback — so the same two minutes of silence came
back on the next surface, and a fresh profile sat at "Translating 0/2"
with nothing else said.

Found only because `e2e:page` runs on a throwaway profile. On a warm one
the pack is already installed and none of this is visible, which is the
profile every hand-check and every quick verification in this session
had been using.

Worth being exact about what is and is not proven: the re-run finished
the PDF in 14 seconds where the previous one timed out at 120. That is
almost certainly a machine-level cache warming between runs rather than
this fix, which reports a download without making one faster. The bug
was real and the fix is right; the number is not evidence for it.

Also gives `e2e:stress` a throwaway profile. It hardcoded one
developer's directory, so it could not run anywhere else and two runs
shared state — the last thing a stress harness should allow. And the
PDF wait in `e2e:page` is seven minutes with elapsed time printed every
minute, because 120 seconds of a first-run download looked exactly like
a hang, and failing to tell those apart has cost three wrong diagnoses
today.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>