test(configurator): CI-gated unit + e2e regression suites, UX/a11y polish - #313

Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.

Inline comments:
In `@configurator/tests-e2e/generator.spec.js`:
- Line 72: The current test assertion in generator.spec.js uses a regex
(/^$|NaN|undefined|null/) that can match substrings; replace it with an anchored
full-string alternation so it only matches entire garbage values. Locate the
expect(...) line in generator.spec.js and change the regex to an anchored
non-capturing group (e.g., /^(?:|NaN|undefined|null)$/) so the .not.toMatch
check rejects exactly empty, "NaN", "undefined", or "null" and not substrings of
valid values.

In `@configurator/tests-e2e/helpers.js`:
- Around line 33-37: The helper readOverrides throws on malformed JSON from
localStorage; update readOverrides to mirror the app's tolerant loader by
wrapping the JSON.parse of localStorage.getItem(STORAGE_KEY) in a try/catch (or
otherwise detect parse errors) and return an empty object {} when parsing fails
or the stored value is invalid, ensuring the function always resolves to a plain
object instead of propagating a SyntaxError.

In `@configurator/tests-e2e/undo-redo.spec.js`:
- Around line 11-53: The test currently uses JSON.stringify(await
readOverrides(page)) which is order-dependent; replace those raw
stringifications with a canonicalized snapshot (e.g., a helper like
canonicalize(obj) that recursively sorts object keys and then JSON.stringify)
and call it wherever you push snapshots and in the undo/redo expect checks
(references: snapshots array, readOverrides(page), the pushes after each action,
and the expect comparisons inside the undo/redo loops). Ensure canonicalize is
used both when building snapshots and when comparing after undo/redo so key
order differences cannot cause flakes.

---

Nitpick comments:
In `@configurator/tests-e2e/shell.spec.js`:
- Around line 61-67: After rapid toggling, the test currently only asserts a
single active side item via the selector '.side__item.side__item--on' which can
miss domain corruption; update the assertion to also verify that the active
item's label remains the expected domain (the one clicked earlier, e.g.
'Borders'). Locate the active element (via
page.locator('.side__item.side__item--on') or by reusing sideItem helper) and
assert its text/content equals 'Borders' (or the expected label) in addition to
ensuring count is 1, keeping the existing watchErrors, gotoClean, and toggle
loop intact.