feat(dictionary): flag unknown words with Add/Block + blocklist screen (#40)

[AsafMah]

Implements #40 (C4-ui) — the clarity layer on top of the #43 blocklist backend.

What it does

• Flag: suggestions that are only learned/typed (source dict = USER_HISTORY/USER_TYPED, i.e. not in a real dictionary) get a subtle underline in the strip. Gated by new pref PREF_FLAG_UNKNOWN_WORDS (default on; toggle in Settings → Dictionaries).
• Add / Block: long-pressing a flagged word opens an AlertDialog with Add to dictionary / Block / Cancel. Curated words keep the existing bin affordance unchanged.
• Blocklist screen: Settings → Dictionaries → Blocklist — per-locale view + remove of blocked words.

Backend (DictionaryFacilitator)

• addToUserDictionary(word) — promote to the personal dictionary and un-blacklist. Resolves the user dict first, so a word is never un-blocked without actually being added.
• blockWord(word) — delegates to removeWord, which (post-fix(dictionary): stop deleted/junk words from resurrecting (incl. via swipe) #43) already permanently blacklists in every group including history-only words; this just gives the UI a clear verb.

Why the source-dict heuristic for flagging

mSourceDict ∈ {USER_HISTORY, USER_TYPED} is cheap (no per-suggestion dict lookup in the view layer) and well-grounded: IMPROVEMENT_PLAN.md already uses mSourceDict != DICTIONARY_USER_TYPED to mean “real dictionary word” — so a valid typed word carries a real-dict source and is not flagged; only typed-but-unknown and learned words are. (#39 C4-smart is the algorithmic follow-up.)

Verification

• Compiles (Java + Kotlin, offline flavor).
• Full :app:testOfflineRunTestsUnitTest: 220 tests, 13 failed — all 13 are the exact dev baseline (4 Windows-only ParserTest + the 9 that PR ci: make unit-test gate green and blocking (#12) #57 quarantines, not yet merged to dev). New DictionaryGroupTest.removeFromBlacklist_unblocksWord passes. Zero new failures.
• Cross-model review (rubber_duck) run; fixed the add-then-unblock ordering bug it surfaced.
• Needs on-device verification (strip underline, dialog, blocklist screen rendering aren't JVM-coverable).

Known limitations (deliberate, for follow-up)

• The flag heuristic can occasionally underline an over-learned word that's also in the main dict (history out-ranks main). The exact predicate (!isInNonHistoryDictionary) exists but needs the facilitator plumbed into the view; deferred to C4-smart: Graduated trust for non-dictionary learned words #39.
• blockWord blocks across all loaded locales (matches the existing bin behavior).
• Blocklist-screen edits write the per-locale file; they take effect on the next dictionary reload (the file is authoritative), not necessarily mid-session.

Implemented coordinator (backend + pref + strings + test) with two parallel subagents (strip UI; settings screen) on disjoint files. Base dev; leaving for review + merge.

[AsafMah]

On-device verified (S24+, standard debug build): unknown-word underline shows, long-press Add/Block dialog works, and the Blocklist settings screen lists/removes. ✅