test(ui): wait for the export's notification, not for the file it wrote

[jimisola]

What & Why

Found on #126 (the setup-uv v10 bump), which had nothing to do with it: pytest (macos-latest, py3.13) went red while py3.12 and py3.14 on the same runner passed, and a plain re-run then went green. Two failures, one root cause, and it is a race that has been in the suite the whole time — it just needs the timing to land badly.

FAILED tests/unit/ui/test_e2e.py::test_demo_e_a_model_that_cannot_classify_still_lists_and_exports
        assert [] == ['Export complete']
FAILED tests/unit/ui/test_models.py::test_the_same_archive_can_be_imported_as_a_separate_copy
        assert (None is not None)      # copy.model_path

The cause

def export_to(page, window, model, dest):
    page.export_selected()
    assert drain_until(window, dest.exists), "export worker never finished"

dest.exists() is satisfied by the worker thread, mid-flight. export_model writes <dest>.tmp and then does an atomic os.replace, so the archive appears the moment the worker is done writing — while _on_exported's notify("Export complete", …) is still on its way back through the bus to the main thread. The wait returns early, with a notification pending.

That breaks two things, in different ways:

1. Directly — the e2e test asserts window.notify.titles on the very next line, one drain too soon, and reads [].
2. At a distance, which is the nastier one — import_from waits for len(window.notify.calls) > before, where before was captured while the export's notification was still queued. The export's own late-arriving notify satisfies that condition, so the helper returns before the import worker has finished. import_model sets model_path during extraction, so the row reads back with model_path still None.

The second failure never mentions exports at all, which is why it looked like two unrelated flakes.

The fix

Wait for the completion that the UI actually reports, and keep the file as an assertion rather than a synchronisation point:

before = window.notify.titles.count("Export complete")
page.export_selected()
assert drain_until(window, lambda: window.notify.titles.count("Export complete") > before), (
    "export worker never finished"
)
assert dest.exists(), "the export reported success without writing the archive"

Three sites: the export_to helper in test_models.py and the two exports in test_e2e.py. import_from is left alone — its wait was the right shape all along, it was just being fed a stale before.

Deliberately not touched: test_models.py's delete test still waits on not paths.model_dir(...).exists(). That worker is run_worker(lambda: shutil.rmtree(...)) with no on_done, so the directory vanishing is the only completion signal there is. Different situation, correctly written.

The regression test

test_export_to_waits_for_the_notification_not_the_archive makes the window deterministic instead of leaving it to which runner you land on: it stubs export_model with one that writes the archive and then takes 250 ms, so the file exists long before the notification does. Against the old helper it fails with IndexError: list index out of range — there is no notification at all yet. Against the new one it passes.

The sleep is simulated slow work inside the export, not the test waiting on an async result; drain_until still does all the waiting, so this doesn't reintroduce what CLAUDE.md §5 rules out.

Test Plan

• pytest — 1448 passed.
• ruff check, ruff format --check, ty check — clean.
• Verified both directions: reverted export_to to the old wait and confirmed the new test fails; restored it and confirmed it passes.

Not reproducible on Linux, which is the point — on this machine drain_until happens to deliver the notification in the same iteration that the file appears. macOS scheduled it differently and CI found it.

Author checklist

• Conventional Commit title
• DCO sign-off
• No CLAUDE.md change — this is a test-helper fix, no architecture or boundary moved