Name a dropped collection language the same way every run (BL-16806)

[andrew-polk]

Problem

Nothing ran the src/BloomE2E suite automatically. Its 14 tests launch a real Bloom.exe, attach
to the WebView2, and click through the workspace tabs, page duplication, and the Text Languages
publish list — but they only ran on the days somebody chose to run them by hand. So a break in the
suite, or a regression in Bloom that the suite would have caught, could sit unnoticed indefinitely.
That is exactly how the component-tester harness sat broken for weeks.

Fix

The nightly workflow now runs BloomE2E as a fourth suite, beside front-end vitest, C# NUnit and
visual regression.

• New Set up BloomE2E tests / Run BloomE2E tests steps, gated on both builds having succeeded
  (the suite launches the Release Bloom.exe, which loads the built front-end assets), with
  !cancelled() so a night that already failed elsewhere still reports this stack.
• It runs before the visual-regression group. Visual regression is the suite that hangs
  (BL-16612) and can consume the job's whole timeout, so anything after it is hostage to a failure
  that has nothing to do with it. Running first means this suite always reports.
• Its own check run, "Nightly tests: BloomE2E (Playwright)", with action_fail_on_inconclusive so
  a suite that reports nothing turns the job red — the trap that once hid the vitest suite
  reporting zero tests.
• Its own run_e2e_tests tick box on a manual run, default on, so the scheduled run is unchanged
  and a developer chasing one suite can pick it alone.
• The testing-inputs fetch is no longer gated on visual regression alone:
  workspace-tabs.spec.ts opens the prepared basic collection.
• On failure only, Playwright's HTML report and traces go up as the e2e-report artifact;
  e2e-junit.xml joins the results artifact and src/BloomE2E/pnpm-lock.yaml the pnpm cache key.
• Everything in the file — tick box, env flag, cache path, steps, publish steps, artifact list,
  header comment — is ordered to match the order the suites now run in.

Two things came out of getting it green:

• The failure-artifact upload could never run. A step if: with no status-check function in it
  gets an implicit success() AND-ed in, and success() is already false once a test step has
  failed — so outcome == 'failure' alone skips the upload on precisely the night it exists for,
  and shows as "skipped" rather than red. The identical pre-existing Upload visual regression screenshots step had the same defect: run 33607591898 failed in that suite and uploaded no
  visual-regression-diffs at all, so the pixel-diff images the BL-16612 hunt wants have never
  been collected. Both steps are fixed.
• The first real CI run diagnosed a known flake. publish-text-languages.spec.ts carried a
  KNOWN FLAKE comment asking whoever saw it fail again to keep the whole log. This run did: of 14
  tests, 12 passed and that one failed on the language name — expected español, got espagnol,
  which is French for Spanish. The collection has just been rewritten to en + fr, so Bloom
  sometimes names the dropped language in the collection's French instead of falling back to the
  autonym. The diagnosis is recorded on the test, which is left running: it is a real
  nondeterminism in what Bloom shows a user, and the only test that catches it.

No playwright install: the fixture attaches to Bloom's WebView2 with chromium.connectOverCDP,
so Playwright launches no browser of its own. PLAYWRIGHT_HTML_OPEN=never keeps the html reporter
from sitting and serving the report at the end of a failing run.

src/BloomE2E/README.md gains an "In CI" section, and the stale sentence in AUTOMATION-DEBT.md
that listed the nightly's suites is corrected.

Verified

A manual nightly run on this branch with only the e2e box ticked
(33665790357) confirmed the
pnpm install, the testing-inputs gate firing without visual regression, all 14 tests discovered and
run against the Release Bloom.exe in 3.9 minutes, the junit file landing where the publish step
reads it, and the check run appearing. That run predates the upload fix, and its
Upload BloomE2E report step duly shows skipped — the bug caught in the act.

Devin review

---

This change is 

[greptile-apps]

Greptile Summary

This PR makes fallback language display names deterministic by using the standard language name while preserving collection-defined names and script/region distinctions.

• Replaces the localized fallback lookup with GetBestLanguageName.
• Explicitly preserves custom names stored in AllLanguages.
• Updates unit and end-to-end expectations to stable standard names.
• Adds regression coverage for metadata-language independence and script/region variants.

Important Files Changed

Filename	Overview
src/BloomExe/Collection/CollectionSettings.cs	Implements deterministic fallback naming and preserves custom and variant display behavior, but silently suppresses all lookup and formatting failures.
src/BloomTests/Book/BookDataTests.cs	Replaces nondeterministic assertions and adds focused coverage for stable names and script/region distinctions.
src/BloomE2E/tests/publish-text-languages.spec.ts	Updates the dropped-language scenario to expect the new stable standard name.
src/BloomExe/Book/BookData.cs	Updates API documentation to describe the deterministic standard-name fallback.

_{Reviews (1): Last reviewed commit: "Keep script and region distinctions on t..." | Re-trigger Greptile}