fix(generator): honor "exclude similar characters" for guaranteed characters

[TheStreamCode]

Summary

The password generator filters look-alike characters out of the pool it draws from, but the one character it guarantees per selected set was picked from the unfiltered alphabet. A generated password could therefore still contain i, l, 1, L, o, 0 or O immediately after the user excluded them.

Worst case, reproduced in a test before the fix: generatePassword(10, { includeNumbers: true, excludeSimilarCharacters: true }) returned 5555555515 — the guaranteed digit was 1, exactly what the option was meant to remove.

The fix filters every alphabet up front and builds the draw pool from the filtered sets, so the guarantee and the pool can no longer disagree. Empty alphabets are dropped instead of sampled (randomInt(0) would throw), and the guaranteed prefix is capped at the requested length so the function cannot return more characters than asked for. That second issue is only reachable through direct API use — the in-app slider starts at 8 and there are at most 4 sets.

No cryptographic behavior changed and no stored data format was touched. Vault encryption, key derivation, storage keys and the KS1 / KS1-PW1 payload layouts are untouched, so existing vaults and previously exported backups are unaffected.

Security review notes

While reviewing the crypto I found trade-offs that are deliberately not changed here, because altering any of them would make existing vaults and exported backups undecryptable. They are now written down in docs/security.md under a new "Known Limitations" section instead of being silently patched:

• Shared AES/HMAC key — KS1 uses the same 256-bit derived key for AES-256-CBC and HMAC-SHA256 rather than splitting it into two subkeys. Encrypt-then-MAC still provides integrity and no practical attack is known, but it does not follow the key-separation principle.
• Backup KDF is weaker than the vault KDF — KS1-PW1 derives its key with PBKDF2 at 100k iterations using the crypto-js default (HMAC-SHA1), not Argon2id. Backups are the one artifact that leaves the device, so this is the weakest link in the current design.
• Six-digit master PIN — a 10^6 keyspace, with no failed-attempt throttling or lockout. Argon2id cost is the only barrier to an offline search by an attacker who extracts both the SecureStore metadata and the encrypted vault. Because each vault stores its own KDF parameters, support for a longer or alphanumeric master password could be added without breaking existing vaults — that is the recommended direction.
• Opt-in screenshot protection, transition-based (not idle-based) auto-lock, Android 13+ clipboard previews, and the trusted OTA update channel.

I attempted to suppress the Android 13+ clipboard preview via the ClipData "is sensitive" extra, but the expo-clipboard bundled with Expo SDK 57 does not expose it through SetStringOptions (neither the flat nor the platform-nested form typechecks). That attempt was reverted and replaced with a comment pointing at the documented limitation; it should be revisited on the next expo-clipboard upgrade.

Cleanup

• Removed src/components/ui/divider.tsx — no importers anywhere in the app or the test suite, and its only styling was a leftover Tailwind class.
• Removed the unused JSX.IntrinsicElements augmentation declaring a style element as any; no JSX <style> element exists and it only weakened typing globally.
• Corrected the nativewind-env.d.ts header: the className augmentations exist for react-native-web CSS hooks (src/utils/webScrollFix.ts), not NativeWind or Tailwind, neither of which is a dependency.
• Added .codex/ to the ignored local agent-state directories and folded the redundant trailing .claude/settings.local.json rule into the .claude/ entry that already covered it (verified both paths are still ignored).

Tests

The generator previously had zero test coverage. Added six regression tests: character-set exclusion (both the multi-set and the single-set case), requested length, short-length behavior, empty option set, and alphabet containment. Three of them fail against the previous implementation and pass against this one — verified by reverting the source and re-running.

Validation

bun run verify — exit 0:

• prettier --check — all matched files use Prettier code style
• tsc --noEmit — clean
• eslint . — clean
• jest --ci --coverage --runInBand — 195 passed, 29 suites (was 189)
• expo-doctor — 20/20 checks passed

bun audit --audit-level=critical (the CI gate) — exit 0, no critical advisories. The unfiltered bun run deps:audit reports 48 pre-existing advisories (33 high, 13 moderate, 2 low), all transitive through the Expo/React Native/Jest/ESLint toolchain (picomatch, js-yaml, lodash, minimatch, node-forge, babel). None affect crypto-js, react-native-argon2, expo-secure-store, expo-clipboard, zod or async-storage. Pre-existing and unchanged by this PR.

Risks

Low. The only behavioral change is in password generation, and it makes the output match what the UI already promised. No migration, no format change, no version bump — the changes are recorded under ## [Unreleased] because app.config.js version mirrors the published Play release by this project'"'"'s documented convention, and no release was cut here.

Checklist

• No vault format, storage key, or KDF parameter change
• No crypto rewritten; findings documented instead
• Full bun run verify green
• Regression tests added and confirmed to fail on the old code
• No secrets added; git ls-files keystore .secrets .env* '"'"'*.jks'"'"' ... returns only the non-secret .env.example
• Changelog updated

🤖 Generated with Claude Code