Build the code under review in the pyinstaller workflow

[happyarts]

The problem

The pyinstaller build smoke test has never run against a pull request.

on: workflow_run starts a job whose actions/checkout resolves to the
repository's default branch, not the commit that triggered the run. So the
workflow has been re-building main every time and reporting success.

Evidence from 2026-07-26 — eight runs, all triggered by pushes to five
different PR branches:

2026-07-26T16:51  workflow_run  branch=main  sha=dc4231e1  success
2026-07-26T16:50  workflow_run  branch=main  sha=dc4231e1  success
2026-07-26T15:24  workflow_run  branch=main  sha=dc4231e1  success
   ... (5 more, same sha)

Same commit every time, and gh pr checks on those PRs lists only
pytest (3.10) and pytest (3.13) — no pyinstaller check to gate on.

Why it matters right now

#330 makes noScribe.main load lazily, which PyInstaller's static analysis
cannot follow. Without a hidden import the packaged app dies at startup with
ModuleNotFoundError while the whole test suite stays green. The
./dist/noScribe/noScribe -h step in this workflow is exactly the check that
catches that class of breakage — it just never saw the branch.

(#330 declares the hidden import, so it is fine; the point is that CI would
not have told either of us.)

The change

Trigger the workflow directly, the same way pytest.yml does. One line, plus
a comment recording why it must not go back to workflow_run.

The trade-off is a build on a PR whose tests are failing. The
"only after pytest succeeded" ordering in the old comment cannot be expressed
across two workflows without giving up the correct checkout — if the build
minutes matter more than the isolation, the alternative is moving the job into
pytest.yml with needs: pytest, which keeps both. Happy to switch to that
if you prefer it.

🤖 Generated with Claude Code