Give the odrcore wrapper typed errors and stop crashing on purpose

[andiwand]

Stacked on #113.

CoreWrapper reported failures through an NSNumber errorCode carrying -2, -3 and -5, decoded again in Document.swift and a third time in DocumentViewController. It also assigned 0 to an object pointer to mean success, so XCTAssertNil(coreWrapper.errorCode) was passing for the wrong reason. It now uses an NSErrorDomain + NS_ERROR_ENUM, which Swift imports as throwing methods.

Crashes fixed

• pagePaths[page] unchecked. The page filter drops every page for some document shapes (a spreadsheet whose only page is named document), so an empty array was reachable and indexing it killed the app. translate now fails with an error rather than reporting success with nothing to show, and the index is clamped.
• backTranslate dereferenced an empty std::optional when called before any translate. Now a typed error — covered by a new test.
• writeContents leaked its DispatchGroup. enter() happened inside the JS completion handler and leave() only on the happy path, so a failing generateDiff left it unbalanced and every save burned the full 30s timeout before throwing. It also called fatalError from that handler. Replaced with a semaphore signalled on every path, plus a main-thread guard so hopping to main can't deadlock.
• Eight fatalError calls used as error handling, for states users actually reach: no document, failed import, missing storyboard identifier, root controller of an unexpected type. Each now recovers or reports.
• Three as! casts on storyboard lookups, and the password alert's textFields![0] / .text!.
• documentEncrypted rescheduled itself when the view wasn't in a window yet — and then presented the alert anyway.

Other cleanups

• DocumentDelegate: class → AnyObject
• @UIApplicationMain → @main
• UIApplication.shared.openURL → open(_:) (deprecated since iOS 10)
• UserDefaults.synchronize() (no-op since iOS 12)
• The progress bar observed UIDocument.progress, not the loadProgress that parse() actually advances — so it never moved.
• The onboarding buttons were hardcoded English in Constants while the app ships 17 localizations. Now NSLocalizedString with intro_next/intro_skip/intro_start added to en.lproj; other languages fall back to English exactly as they do today, but are now translatable via Crowdin. The rest of Constants was dead code.

Verification

• Both flavors build; both install and launch in a simulator and stay running.
• Test suite goes from 1 test (a measure block asserting a nil error code) to 5, all passing: page output, back translation, backTranslate before translate, unsupported file type, and the performance measurement.

Deliberately not in this PR

The actor + async/await restructure of CoreWrapper. parse() still runs the C++ translate synchronously on whatever thread the didSet fired on. That change alters threading semantics, and with no integration coverage over document loading I would be changing behavior I cannot verify. It wants its own PR with tests written first — say the word and I will do it next.

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: a50c161555

ℹ️ About Codex in GitHub

Codex has been enabled to automatically review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

When you sign up for Codex through ChatGPT, Codex can also answer questions or update the PR, like "@codex address that feedback".

[andiwand]

Correct, fixed in 30864e6.

A key missing from a table resolves to the key itself rather than falling back to the development language, so the other 16 localizations would have rendered intro_next / intro_start on the button where they previously showed the hardcoded English.

Went with the explicit-value option rather than seeding 16 tables with untranslated English:

NSLocalizedString("intro_start", value: "Start", comment: "onboarding button on the last page")

Verified the fallback actually resolves to Start / Next and not the key when no matching table exists. Translations can land per-locale later without touching the call site.

One loose end from the same commit, out of scope here: intro_skip was added to en.lproj but nothing reads it — the Skip button's title is still hardcoded in Main.storyboard and has no outlet.